Paper deep dive
IntrAgent: An LLM Agent for Content-Grounded Information Retrieval through Literature Review
Fengbo Ma, Zixin Rao, Xiaoting Li, Zhetao Chen, Hongyue Sun, Yiping Zhao, Xianyan Chen, Zhen Xiang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 6/21/2026, 5:44:03 AM
Summary
The paper introduces 'IntraView', a new task for fine-grained, content-grounded information retrieval from scientific literature, and 'IntrAgent', an LLM-based agent designed to solve it. IntrAgent mimics human reading behavior through a two-stage pipeline: a 'Section Ranking' stage that uses structural-knowledge-enabled reasoning to prioritize relevant sections, and an 'Iterative Reading' stage that extracts details and performs a sufficiency check to prevent hallucinations. To evaluate this task, the authors propose 'IntraBench', a benchmark consisting of 315 test instances across five STEM domains (physics, earth science, public health, engineering, and material science). Experimental results show that IntrAgent outperforms state-of-the-art RAG and research-agent baselines in cross-domain accuracy.
Entities (7)
Relation Signals (5)
IntrAgent β addresses β IntraView
confidence 100% Β· propose IntrAgent, an LLM-based agent that addresses this challenging task [IntraView].
IntrAgent β consistsof β Section Ranking
confidence 100% Β· It follows a two-stage pipeline: a Section Ranking stage... and an Iterative Reading stage...
IntrAgent β consistsof β Iterative Reading
confidence 100% Β· It follows a two-stage pipeline: a Section Ranking stage... and an Iterative Reading stage...
IntraBench β evaluates β IntraView
confidence 100% Β· To support rigorous evaluation, we introduce IntraBench, a new benchmark... for evaluating IntraView approaches.
IntrAgent β outperforms β RAG
confidence 90% Β· IntrAgent achieves on average 13.2% higher cross-domain accuracy than state-of-the-art RAG...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Scientific research relies on accurate information retrieval from literature to support analytical decisions. In this work, we introduce a new task, INformation reTRieval through literAture reVIEW (IntraView), which aims to automate fine-grained information retrieval faithfully grounded in the provided content in response to research-driven queries, and propose IntrAgent, an LLM-based agent that addresses this challenging task. In particular, IntrAgent is designed to mimic human behaviors when reading literature for information retrieval -- identifying relevant sections and then iteratively extracting key details to refine the retrieved information. It follows a two-stage pipeline: a Section Ranking stage that prioritizes relevant literature sections through structural-knowledge-enabled reasoning, and an Iterative Reading stage that continuously extracts details and synthesizes them into concise, contextually grounded answers. To support rigorous evaluation, we introduce IntraBench, a new benchmark consisting of 315 test instances built from expert-authored questions paired with literature spanning five STEM domains. Across seven backbone LLMs, IntrAgent achieves on average 13.2% higher cross-domain accuracy than state-of-the-art RAG and research-agent baselines.
Tags
Links
- Source: https://arxiv.org/abs/2604.22861v1
- Canonical: https://arxiv.org/abs/2604.22861v1
Trouble viewing inline? Open PDF directly β
Full Text
161,586 characters extracted from source content.
Expand or collapse full text
IntrAgent: An LLM Agent for Content-Grounded Information Retrieval through Literature Review Fengbo Ma 1,* , Zixin Rao 1,* , Xiaoting Li 1 , Zhetao Chen 1 , Hongyue Sun 1 , Yiping Zhao 1 , Xianyan Chen 1,β , Zhen Xiang 1,β 1 University of Georgia, Athens, GA, USA * Equal contribution; β Corresponding authors: xychen@uga.edu, zxiangaa@uga.edu Abstract Scientific research relies on accurate informa- tion retrieval from literature to support ana- lytical decisions. In this work, we introduce a new task, INformation reTRieval through literAture reVIEW (IntraView), which aims to automate fine-grained information retrieval faithfully grounded in the provided content in response to research-driven queries, and pro- poseIntrAgent, an LLM-based agent that addresses this challenging task. In particu- lar,IntrAgentis designed to mimic human behaviors when reading literature for infor- mation retrieval β identifying relevant sec- tions and then iteratively extracting key de- tails to refine the retrieved information. It follows a two-stage pipeline: a Section Rank- ing stage that prioritizes relevant literature sections through structural-knowledge-enabled reasoning, and an Iterative Reading stage that continuously extracts details and synthesizes them into concise, contextually grounded an- swers. To support rigorous evaluation, we in- troduceIntraBench, a new benchmark con- sisting of 315 test instances built from expert- authored questions paired with literature span- ning five STEM domains. Across seven back- bone LLMs,IntrAgentachieves on average 13.2% higher cross-domain accuracy than state- of-the-art RAG and research-agent baselines. 1 Introduction In modern scientific research, accurately extract- ing metadata, experimental setups, and contextual knowledge from prior literature is essential. However, such information retrieval is chal- lenging due to the complexity of scientific liter- ature (Huang et al., 2025), which often requires deep domain expertise and substantial time to read and interpret (Wu et al., 2025; Lu et al., 2024a; Li IntrAgentProject page:https://intragent.github. io/ .IntrAgentCode:https://github.com/FengboMa/ IntrAgent.IntraBenchDataset:https://huggingface. co/datasets/IntrAgent/IntraBench. et al., 2024; Wang et al., 2024c). Therefore, an automated system for researchers to accurately and efficiently extract relevant information from liter- ature can potentially transform research practices across many scientific disciplines. Motivated by this need, we propose a new task called INformation reTRieval through literAture reVIEW (IntraView) β given a literature and an information retrieval query, extract and synthesize information faithfully grounded in the provided content. WhileIntraViewfollows the general for- mat of Content Question Answering (CQA) (Jin et al., 2019; Mihaylov et al., 2018; Welbl et al., 2017; Chen et al., 2021; Reddy et al., 2019; Chen et al., 2017), it introduces greater challenges due to the structural complexity and domain-specific language of scientific literature, in contrast to the more generic knowledge chunks typically used for standard CQA (Dasigi et al., 2021; Lu et al., 2022; Joshi et al., 2017; He et al., 2024; Rajpurkar et al., 2018; Rein et al., 2023; Jin et al., 2021; Krithara et al., 2023). Moreover, in practical applications ofIntraView, it is crucial to avoid hallucination by explicitly acknowledging when the requested information is not present in the literature. Due to these critical differences, existing methods for conventional CQA tasks are insufficient for ad- dressing the new challenges ofIntraView(Wu et al., 2024b; Lu et al., 2024c; Wang et al., 2023a). Direct LLM querying for specific information re- trieval β even with models fine-tuned for scien- tific domains (Zhang et al., 2024c) β struggles to manage the overwhelming information contained in a full literature. Retrieval-augmented gener- ation (RAG) approaches attempt to address this limitation by selecting a small set of text chunks deemed most relevant to the query (Lewis et al., 2020; Wang et al., 2024d; Ye et al., 2024), typ- ically based on semantic similarity, and then us- ing these chunks to prompt the LLM. However, such methods rely solely on surface-level semantic 1 arXiv:2604.22861v1 [cs.IR] 23 Apr 2026 similarity, often failing to align domain-specific terminology in the query with the actual relevant content, and ignore the rich structural organiza- tion inherent to scientific documents. Recently, numerous LLM-based agents have been developed to tackle complex tasks (Bran et al., 2024; Yin et al., 2024; Wu et al., 2024a; Tang et al., 2023; Nguyen et al., 2025), including those involving scientific literature, such as literature search, summarization, research idea exploration, and academic writing assistance. Compared to standalone LLMs, these agents leverage the reasoning capabilities of LLMs for comprehensive task planning and interact with external tools and environments to enhance task execution (Whitfield et al., 2023). However, the scientific QA task addressed by these agents fun- damentally differs from ourIntraView. They aim to answer general scientific questions through ex- ternal resource exploration, whereasIntraView focuses on retrieving and reasoning strictly con- strained to a provided paper. Thus, these methods perform poorly onIntraView, as will be demon- strated in our experiments. To close this gap, we propose an LLM agent, IntrAgent, as the first specialized solution to IntraView. The key idea behindIntrAgentis to emulate human behavior when reading techni- cal literature for information retrieval β identifying the most relevant sections and progressively accu- mulating details until the query is fully addressed (Suppawittaya, 2021; Miller, 1956). Accordingly, IntrAgentcomprises two main stages. First, in the section ranking stage, paper sections are reordered based on structure-aware reasoning by the LLM to reflect their relevance to the query. Second, in the iterative reading stage, the reordered sections are sequentially accessed to extract relevant de- tails, continuing until the accumulated information is deemed, via LLM reasoning, sufficient to an- swer the query. Notably,IntrAgentincorporates several novel designs to address the unique charac- teristics and challenges ofIntraView. For exam- ple, we introduce a hierarchy preservation step that captures the structural organization of the research literature, enabling more comprehensive reason- ing during section ranking. Additionally, during iterative reading, we design a sufficiency check mechanism to assess whether the extracted details are adequate for answering the query, ensuring the faithfulness by reducing the risk of hallucination. In addition, we introduceIntraBench, the first benchmark designed to evaluateIntraViewap- proaches, including our proposedIntrAgent. The benchmark consists of 315 test instances drawn from five scientifically and societally significant domainsβphysics, earth science, public health, en- gineering, and material science. We further assess IntrAgentagainst state-of-the-art RAG systems and literature-oriented agents, demonstrating its superior performance across all domains. Our key contributions are summarized below: β’We introduceIntraView, a novel task for accu- rate, automated, and content-grounded informa- tion retrieval from a provided scientific literature, and proposeIntrAgent, an LLM agent specifi- cally designed to tackle this task. β’We develop a novel two-stage pipeline for IntrAgent, consisting of section ranking and iterative reading, designed to mimic human read- ing behavior. We introduce a hierarchy preserva- tion mechanism to help the agent leverage struc- tural knowledge for more effective section rank- ing, and implement a sufficiency check to miti- gate hallucination during iterative reading. β’We proposeIntraBench, the first benchmark for evaluatingIntraView, consisting of 315 test in- stances across five impactful domains. β’We evaluateIntrAgentonIntraBenchand show that it outperforms both state-of-the-art RAG and literature-agent baselines across rep- resentative backbone LLMs in average cross- domain accuracy. 2 Related Work Retrieval-Augmented Generation (RAG)RAG enhances LLMs by retrieving relevant documents for response generation (Gao et al., 2024), enabling external knowledge integration in knowledge- intensive tasks (Wu et al., 2025). The vanilla RAG framework adopts a three-stage pipeline consisting of indexing, retrieval, and generation (Lewis et al., 2020), but it faces challenges such as retrieval noise, hallucinated outputs, and limited reasoning over re- trieved content (Zhu et al., 2024). Recent RAG variants improve embedding quality, retrieval accu- racy, and context control through techniques such as re-ranking (Ye et al., 2024), dynamic embed- dings (Jiang et al., 2024), contextual retrieval (An- thropic, 2024), external memory (Li et al., 2024; Wang et al., 2024c), and modular pipelines (Lu et al., 2024a). However, most of these approaches still follow a flat retrieve-then-generate architecture, 2 which limits their effectiveness for fine-grained in- formation retrieval from structured scientific con- tent, as required by IntraView. LLM Agent for Literature TasksMany domain- specific research agents have been developed across scientific disciplines, including chemistry (Bran et al., 2024; Tang et al., 2025), engineering (Zhang et al., 2024b; Singh et al., 2025), mathematics (Wu et al., 2024b), and bioinformatics (Qi et al., 2024); but they are tailored to tasks within a single field. There also exist agents for cross-domain literature- related tasks other thanIntraView, such as lit- erature search (Agarwal et al., 2024), research ideation (Lu et al., 2024b), and end-to-end scien- tific writing (Schmidgall et al., 2025; Shao et al., 2024). However, most of these agents are not di- rectly applicable toIntraView, which presents dis- tinct goals and requirements. While some agents for general scientific QA via literature search, such as PaperQA (LΓ‘la et al., 2023), PaperQA2 (Skarlinski et al., 2024), QASA (Lee et al., 2023), and SciMaster (Chai et al., 2025), can be adapted toIntraViewby disabling their retrieval component and supplying the same paper as input, these systems are not designed for IntraViewand therefore cannot match the perfor- mance of ourIntrAgent, as will be demonstrated in our extensive evaluation. 3 Proposed IntraView Task Information retrieval from scientific literature is critical to a wide range of subjects. The retrieved in- formation could be used to guide the experimental design, hypothesis refinement, simulation configu- rations, statistical methodology, results validation, and other downstream decision-making in research workflows (Rothstein et al., 2024; Kumar et al., 2024; Yadav et al., 2022; Senapati et al., 2024). In this work, we propose a novel task, IN- formation reTRieval through literAture reVIEW (IntraView), which aims to accurately extract key information from a provided research paper in re- sponse to a specific query. For example, given a Surface-enhanced Raman spectroscopy (SERS) pa- per with query βWhat is the excitation laser wave- length used for SERS measurements?β one should respond with a specific laser wavelength solely based on the experimental context described in the paper (Zhao et al., 2024). Formally, we construct IntraViewas a CQA problem (Jin et al., 2019). The objective is to build an automated system that, when given a literatureCand a research-driven questionQ, generates an accurate and content- grounded answerAby identifying the most rel- evant information in C without hallucination. Compared to existing CQA tasks,IntraView differs in two aspects: (a) it provides the full litera- ture rather than a pre-selected or processed chunk, with the relevant information potentially appear- ing anywhere in the literature or not at all; and (b) it tackles domain-specific queries that may re- quire cross-referencing multiple sections beyond where the final answer resides. Compared with other literature-related tasks, such as scientific QA via literature search,IntraViewemphasizes faithful information retrieval restricted to the pro- vided content, independent of the external validity of its scientific claims. In contrast to other ex- isting tasks such as PeerQA(?), where answers may involve author intent, argumentative reason- ing, or external domain knowledge beyond the pa- per,IntraViewstrictly requires that all answers be directly grounded in the provided content, ne- cessitating structured, multi-stage reading rather than generative reconstruction. Therefore, existing approaches for those tasks cannot effectively tackle IntraView, as will be shown by our extensive eval- uation. 4 Proposed IntrAgent Framework 4.1 Overview of IntrAgent As the first agent framework specially designed for IntraView,IntrAgentemploys βmindset bionicsβ approach to emulate the natural reading workflow of humans during information retrieval β it begins with a guiding question, infers the section most likely to contain the answer, extracts the key infor- mation, evaluates whether the question has been sufficiently addressed, and iterates through addi- tional sections as needed (Miller, 1956; Suppawit- taya, 2021). Accordingly,IntrAgentcomprises two major stages β section ranking (Section 4.2) and iterative reading (Section 4.3) β as illustrated in Figure 1. In the section ranking stage, the agent identifies the most relevant sections of a literature using an LLM by structure-aware reasoning. In the iterative reading stage, it repeatedly gathers information and evaluates sufficiency until the input query can be adequately answered. Our design ofIntrAgentoffers several advan- tages over approaches for other CQA tasks, such as 3 RAG (Lewis et al., 2020), making it well-suited for IntraView: 1) Effective context prioritizing: Un- like RAGβs semantic-similarity-based chunk rank- ing, our reasoning-based section ranking more precisely locates the details relevant to the query within the literature. 2) Explicit hallucination miti- gation: The sufficiency check in the iterative read- ing stage determines whether additional reading is necessary, thereby explicitly reducing hallucination by ensuring that only substantiated answers are re- turned. A toy example ofIntrAgenthandling a SERS-related query is shown in Figure 2. 4.2 Section Ranking Scientific research literature typically follow a well- defined section hierarchy: parent-level headings convey broader topics, while sub-level headings provide more specific details. Unlike flat, semantic- similarity-based RAG that overlooks document structure and often fails to align a scientific ques- tion with the relevant sections (Lewis et al., 2020), IntrAgentleverages this structural knowledge for reasoning-based ranking, prioritizing sections rele- vant to the question through three key steps: Section Heading Parsing To standardize struc- tural information across literature with diverse tem- plates, we convert each input literatureCinto a Markdown-formatted versionC β² with systematic markers for section headings. For papers provided as PDFs, which is the most common case, we use minerU (Wang et al., 2024a), a visual model for layout and section detection, for the conversion. Hierarchy Preservation This step aims to con- struct a section tree that represents the hierarchical structure of the literature. This tree will support: 1) LLM-based reasoning for section ranking, and 2) structured text parsing for iterative reading. We begin by extracting all section and subsection head- ings from the Markdown-parsed textC β² to form an initial heading setH 0 1 . Next, we prompt an LLM to infer the hierarchical relationships among these headings, treating each parent section as a node with child nodes corresponding to its subsec- tions. The prompt is detailed in Appendix E.1.1. The LLM returns a set of paths from the root to each node in the hierarchy. To eliminate redundant nodes, we remove paths where a parent section is immediately followed by a subsection without intervening content. 1 Markdown files from minerU uniformly prepend a single pound sign to all section and subsection titles. The resulting filtered set of headings is denoted asH =h 1 ,h 2 ,...,h n . Reasoning-Based Ranking Given the filtered heading setHand the research questionQ, the model is prompted to reason about which section is most likely to contain the informa- tion needed to answerQβ this design mirrors nat- ural human reading behavior (see Appendix E.1.2 for the detailed prompt). The output of this step is a ranked list of indices R = [r 1 ,...,r n ]. UsingR, we reorder the sec- tions to construct a list of (heading, text) pairs C R = [(h r 1 ,t r 1 ),..., (h r n ,t r n )] , wheret r i de- notes the text in the literature corresponding to sec- tionh r i . These reordered sections will be sequen- tially accessed during the iterative reading stage described next. 4.3 Iterative Reading At each step, the agent selects its next action from a predefined set: reordered section access, section detail extraction, and information sufficiency check, based on reasoning over the outcomes of the previ- ous step. This action space is carefully designed to enable targeted retrieval of key information relevant to the input query while suppressing hallucination. Reordered Section Access When reading is initiated or a section(h r i ,t r i )has been fully processed, the agent retrieves the next section (h r i+1 ,t r i+1 )from the listC R . This ensures that sections are read in descending order of estimated relevance to the research questionQ, maintaining a focused and efficient reading trajectory. Section Detail Extraction The agent examines the current section(h r i ,t r i ) to extract information relevant toQ. This action is designed to identify and record key scientific detailsD i , including terminology, numerical data, experimental results, measurements, statistical indi- cators, conclusions, and any comparative or causal statements that explicitly address the research ques- tion. Each detail inD i is anchored to its original sentence and stored in a short-term memory for final answer synthesis. See Appendix E.2.3 for detailed prompts. Information Sufficiency Check This step gov- erns the action loop of iterative reading by de- termining whether more sections need to be ac- cessed. The agent uses an LLM to reason over the detailsD i gathered in the current iteration and assess whether they are sufficient to answer the research questionQ. The reading loop continues 4 Action Loop w Research Question Q Research Paper C Hierarchy Preservation Final Answer Synthesis Output Information Sufficiency Check Section Detail Extraction β β Reordered Section Access Reasoning-based Ranking & Reordering Section Title Parsing β: Gathered Information is insufficient β: Gathered Information is sufficient Input Section Ranking : LLM : Input: Deterministic Python code : Steps involving LLM Format Conversion Iterative Reading Figure 1: Overview of theIntrAgentpipeline containing two stages: Section Ranking (top) reorders the paperβs sections by relevance to the Research QuestionQ, while Iterative Reading (bottom) steps through ranked sections, extracting information until gathered information is sufficient. with the next section if the LLM outputsNO; oth- erwise, it terminates ifYESis returned. Notably, our prompt here includes explicit instructions to avoid speculation and hallucination, as detailed in Appendix E.2.4. The relevant evidence for a given query may be distributed across multiple non-adjacent sections of a paper. The sufficiency check explicitly prevents premature termination by requiring the agent to continue reading until suffi- cient supporting information has been accumulated across sections, enabling reliable cross-section syn- thesis. Moreover, we introduce a confidence-based mechanism that supports three reading styles: con- servative, balanced (default), and aggressive, al- lowing the user to control the operational over- head β conservative reading accesses fewer sec- tions, while aggressive reading explores more. Rel- evant prompts are detailed in Appendix E.4. Once the sufficiency check terminates the ac- tion loop after readingmsections, the agent in- vokes an LLM to synthesize the final answerA = LLM(D 1 ,...,D m ,Q)using the accumulated de- tailsD 1 ,...,D m during the iterative reading. Detailed prompts are provided in Appendix E.2.5. 5 Proposed IntraBench for Evaluation 5.1 Overview of IntraBench As a novel task,IntraViewlacks dedicated evalu- ation benchmarks. Thus, we proposeIntraBench, a benchmark specially designed forIntraView, which demands expert-level understanding of domain-specific contexts.IntraBenchconsists of 315 test instances, derived from expert-curated insights and grounded theory (Hallberg, 2010) β paired with research literature from five highly im- pactful domains spanning physics, earth science, Benchmark# InstancesDomains LitQA50Bio LitQA2248Bio IntraBench315Phys, ES, PH, Engr, MS Table 1: ComparingIntraBenchwith existing bench- marks for scientific QA via literature search (Phys: Physics, ES: Earth Science, PH: Public Health, Engr: Engineering, MS: Material Science, Bio: Biology). public health, engineering, and material science. Although focusing on different tasks,IntraBench contains more test instances from broader scientific fields compared with existing benchmarks for QA via literature search (see Table 1). More details on benchmark construction, including the domain choices, are deferred to Appendix A. 5.2 LLM-Grounded Multiple-Choice Evaluation for IntraBench In real-world applications,IntraViewis expected to produce concise, free-form answers rather than selecting from predefined choices. However, this presents two major challenges for evaluation. 1) Scientific terminologies often involve abbreviations and synonyms β for example, terms like AgNP, sil- ver nanoparticle, and silver nanorod may all refer to the same surface-enhanced substrate treatment in a SERS experiment. 2) When retrieved information is numerical or factual, precise accuracy is criti- cal. Traditional string evaluation methods, such as BLEU (Papineni et al., 2002) or ROUGE (Lin, 2004), are unsuitable since they rely on surface- level text similarity and often fail to capture seman- tic correctness in specialized scientific contexts. To address these challenges, we adopt a multiple-choice format when creating questions forIntraBench, though the evaluated methods 5 must still generate short answers without being shown the multiple-choice options. During evalua- tion, an LLM maps each generated short answer to the most relevant multiple-choice candidate. This relevance-based mapping effectively addresses the first challenge of varied terminologies such as ab- breviations, synonyms, or domain-specific expres- sions referring to the same concept. If the LLM cannot confidently align the generated answer with any provided choice (due to insufficient relevance), we use a fallback label β βNone of the aboveβ β to avoid misclassification of ambiguous responses. The final evaluation metric is accuracy based on the multiple-choice mapping. Empirically, this method demonstrates strong alignment (Average correctness agreement of 63/65 on the Physics dataset using GPT-4.1) with manual mappings by domain experts; see Section 6.4 for details. 5.3 Construction of IntraBench Dataset Selection of PapersFor each domain, papers are manually selected by a corresponding expert on our team. To reduce selection bias, each expert first curates a larger pool of familiar papers, from which five are randomly chosen. The selected papers are constrained to impactful, peer-reviewed journals to ensure their authority. Expertsβ familiarity with these papers ensures accurate annotation of the ground-truth answers for the associated questions. Creation of Questions We aim to capture both technical depth and conceptual complexity through expert-level inquiry. To this end, the questions are generated by our domain experts based on their natural reading practices, i.e., what information they would seek from a paper within their field. These questions are categorized into four task- oriented categories based on general research prin- ciples (Dillon, 1984): study subject & experimental setup, data characteristics & collection, technical approach & details, conclusions & results. Details about question setup in Appendix A.3 Creation of Answer ChoicesFor each question and its paired paper, six answer options are created by domain experts, including one correct answer and five distractors. The six options consistently in- clude βAll of the aboveβ and βNone of the aboveβ, either of which may serve as the correct answer when applicable. Distractors are manually con- structed based on: 1) concepts, numerical values, or textual information from the paper that closely resemble the correct answer; or 2) commonly used information or conventions in the respective field. 6 Experiments Our experiments aim to address the following re- search questions: (1) RQ1 (Section 6.2): Can IntrAgenteffectively solve theIntraViewtask compared with the baselines?(2) RQ2 (Sec- tion 6.3):Do our designed components for IntrAgent, such as hierarchy preservation, the confidence level mechanism, and the information sufficiency check, function as intended? (3) RQ3 (Section 6.4): IsIntrAgentrobust to variations in evaluation and input conditions, such as different mapping models and non-standard headings? 6.1 Experiment Settings Dataset and Evaluation Metrics We use our proposedIntraBenchfor evaluation. Following Section 5.2, we use GPT-4.1 to map each short answer generated by the method to one of the pre- defined answer choices. Evaluation results using alternative mapping models are presented in Sec- tion 6.3. Our primary evaluation metric is accuracy after the answer mapping. Baseline We evaluate a broad range of retrieval- augmented generation (RAG)βbased approaches widely adopted for CQA and scientific reason- ing, including (1) vanilla RAG using embedding models all-MiniLM-L6-v2 (Wang et al., 2020), E5-mistral-7b-instruct (Wang et al., 2023b), and GritLM-7B (Muennighoff et al., 2024), respectively, with cosine-similarity retrieval over 500-token chunks (50-token overlap); (2) contextual RAG variants that enhance retrieval via dynamic chunk selection and adaptive context expansion (An- thropic, 2024); and (3) advanced RAG extensions such as DRAGIN (Su et al., 2024b), RΒ²AG (Ye et al., 2024), and LongRAG (Jiang et al., 2024), which introduce multi-hop reasoning, re-ranking, and long-context retrieval. We also consider representative literature- focused agents, including LUMOS (Yin et al., 2024), PaperQA2 (Skarlinski et al., 2024), Agentic- Hybrid-RAG (Nagori et al., 2025), and SciMas- ter (Chai et al., 2025), the last of which reports a score of 32.1 on Humanityβs Last Exam, marking the state of the art of scientific agents. All baselines follow the sameIntraBencheval- uation protocol described above. 6 Default hyperparameters are used unless other- wise specified; detailed configurations are provided in Appendix F. 6.2 Main Result Table 2 show thatIntrAgentsets a new state of the art across all five domains ofIntraBenchfor seven backbone LLMs. Averaged over physics, public health, earth science, engineering, and mate- rial science,IntrAgentachieves 70.0% with GPT- 4o, 75.8% with GPT-4.1, 74.4% with DeepSeek- R1, 73.4% with o3, 73.8% with o4-mini, 75.9% with Gemini-2.5 Pro, and 68.8% with Llama-3.1- 70B. Compared with the strongest baseline under each model,IntrAgentsurpasses them by 7.9%, 11.6%, 8.9%, 18.2%, 17.4%, 21.0%, and 7.4%, re- spectively. These results suggest that performance gains stem not from increased context length, but from targeted evidence selection: unlike chunk- based methods that introduce irrelevant context, IntrAgent isolates task-relevant sections, enabling more precise reasoning. Such a performance gain stems from our novel agent design: 1) hierarchy-aware section ranking based on reasoning, and 2) sufficiency check that halts reading once evidence is complete. In contrast, RAG sequentially feeds multiple un- structured text chunks into the LLM, which often introduces irrelevant or noisy information. The agent baselines, originally designed for scientific QA through online search, degenerate into static in- formation retrieval pipelines similar to RAG when they are directly provided with the literature for IntraView. A representative example demonstrating the ad- vantage ofIntrAgentover the RAG baseline is shown in Figure 2, along with the reasoning tra- jectory forIntrAgentβs section ranking and suffi- ciency check. In this example, the agent prioritizes the Experiment Setup β SERS measurement sec- tion based on the reasoning that laser wavelength information is typically part of the experimental setup. During the information sufficiency check, this section is verified to explicitly state the exci- tation wavelength (785 nm), confirming that the relevant detail is correctly retrieved from the ap- propriate section. These results address RQ1, con- firming thatIntrAgentmore effectively solves the IntraView task compared to the baselines. 6.3 Ablation Studies Study on Hierarchy Preservation We evaluate IntrAgentwithout the hierarchy preservation step by directly ranking the raw section headingsH 0 ex- tracted from the Markdown-parsed text. As shown in Table 3, removing the hierarchy preservation step leads to a clear drop in cross-domain accuracy for all three model choices, highlighting the neces- sity of this key step for effectively incorporating structural context into section ranking. Study on Confidence LevelIntrAgentsup- ports three confidence levels via prompt variation during iterative reading: 1) Conservative, which halts only when all answer components are ex- plicitly observed; 2) Balanced (the default), which stops when evidence appears sufficient; and 3) Ag- gressive, which allows early termination based on partial evidence. The full prompts are provided in Appendix E.4. This confidence level parameter en- ables users ofIntrAgentto control the operational overhead (more details on overhead are shown in Appendix B.3). As shown in Table 4, more aggres- sive reading reduces the number of iterative reading steps but at the cost of lower information retrieval accuracy. Interestingly, the conservative mode, de- spite reading more sections, performs the worst, consistent with RAG observations that performance degrades with very long context windows (Jiang et al., 2024; Yepes et al., 2024; Wang et al., 2024b). Study on Information Sufficiency Check We evaluateIntrAgentwithout the information suffi- ciency check by accessing only the top-1 section during iterative reading. On the physics dataset, using GPT-4o as the backbone LLM and GPT-4.1 as the mapping model, this modification results in a substantial accuracy drop from 75.4% to 32.2%, showing the necessity of this critical component. Moreover, analysis of failure cases reveals two key issues: 1) Incomplete retrieval: The necessary in- formation could be distributed across multiple sec- tions or require cross-referencing. Although the most relevant section is correctly identified in most cases, it alone may not suffice to answer the query β leading the agent to respond with βNone of the above.β 2) Hallucination: The agent produces confi- dent but unsupported answers. An example failure case for hallucination is provided in Appendix C.1. These results confirm that all the designed com- ponents function as intended, addressing RQ2. 7 MethodGPT-4oGPT-4.1DS-R1o3o4-miniGemini-2.5 ProLlama-3.1-70B RAGVanilla RAG all-MiniLM-L6-v260.361.264.360.461.561.859.2 Vanilla RAG E5-mistral-7b-instruct59.464.263.860.361.459.960.5 Vanilla RAG GritLM-7B60.463.263.259.758.458.461.4 Context. RAG E5-mistral-7b-instruct60.763.862.859.158.358.958.9 Context. RAG GritLM-7B60.862.861.658.460.761.659.2 DRAGIN42.544.646.944.046.945.945.4 R 2 AG59.459.561.556.655.355.656.1 LongRAG62.164.765.557.058.357.157.4 AgentLUMOS50.252.155.455.256.454.954.4 PaperQA247.748.954.051.849.251.253.8 Agentic-Hybrid-RAG59.860.262.357.557.857.256.6 SciMaster59.057.663.357.258.157.257.0 IntrAgent (Ours)70.075.874.473.473.875.968.8 Table 2: Cross-domain accuracy (in %, defined by the macro average over the five domains) onIntraBench. Our IntrAgentuniformly outperforms the RAG-based retrieval and agent-based baselines across the five domains for seven model choices. See Appendix B.1 for the complete breakdown results. Input Q: What is the excitation laser wavelength used for SERS measurements? RAG Section Heading Parsing: 1. Abstract; 2. Introduction ... 5. Experiment Setup, 6.SERS measurement ...12. Conclusion Hierarchy Preservation: 1. Abstract; 2. Introduction ... 5. Experiment Setup; 6. Experiment Setup - SERS measurement ...12. Conclusion Reasoning-Based Ranking: Rank1: Experiment Setup - SERS measurement; Rank 2: Introduction; Rank 3: Conclusion ... Rank 11...; Rank 12... Reordered Section Access: 6. Experiment Setup - SERS measurement + <text> Section Detail Extraction: <details> ... 785 nm laser ... Information Sufficiency Check: β YES Answer: The laser wavelength used is 785 nm;... as stated in the sentence ... β C: <PDF> Machine Learning Enabled Multiplex Detection of Periodontal Pathogens by Surface-enhanced Raman Spectroscopy β IntrAgent Extracted Chunk: ... among others, leading to more intricate Ramanspectra ... attainment of SERS at high fidelity...measurements were conducted ... Answer: ... The evaluation entries do not specify the laser wavelength ... Agent Reasoning for Section Ranking: Rank 1: 6. Experiment Setup - SERS measurement β This section is most likely to contain information about the laser wavelength as part of the experiment setup. Rank 2: 2. Introduction β The introduction often provides some context of the experiment. ... Output: Rank1: Experiment Setup - SERS measurement; Rank 2: Introduction; Rank 3: Conclusion ... Rank 11...; Rank 12... Agent Reasoning for Information Sufficiency Check: This is Section 6. Experiment Setup - SERS measurement. This section clearly stated experiment conditions including laser wavelength, also other experiment conditions including ... Output: Sufficiency of the chunk β YES Figure 2: An example ofIntrAgentexecuting a question-paper pair fromIntraBench. For an input questionQ regarding paperC, vanilla RAG fails to extract the correct chunk, resulting in an incorrect answer. In contrast, IntrAgentranks the sections through reasoning and retrieves the correct details that pass the sufficiency check in the first iteration, leading to a correct answer. Details for section ranking and sufficiency check are also presented. ModelGPT-4o GPT-4.1 DeepSeek-R1 w/o HP60.770.364.2 w/ HP65.672.567.6 Table 3: Accuracy (%) ofIntrAgentw/ and w/o the hi- erarchy preservation (HP) step for three model choices. 6.4 Robustness of IntrAgent to Mapping Model and Input Variability Impact of Mapping ModelSince the evaluation protocol ofIntraBenchinvolves mapping short- form answers to multiple-choice (MCQ) selections, it is important to assess the reliability of the map- ping model. Using identical prompts on the physics sub- set, we compare the mappings produced by GPT- MetricConservative Balanced Aggressive Accuracy (%)58.968.362.7 Avg. Iterations9.95.13.9 Med. Iterations1121 Std Dev. Iterations7.95.55.3 Med. Token Count785363762233 Table 4: Impact of confidence level on the MCQ accu- racy, the number of reading iterations, and token count. 4o, GPT-4.1, and DeepSeek-R1 against ground- truth annotations provided by domain experts. Here, we consider short-form answers generated by IntrAgent(with GPT-4o as the backbone) and by vanilla RAG. We also define a correctness agree- ment metric as the ratio of MCQ choices that match the human label in correctness. As shown in Table 5, GPT-4o and GPT-4.1 yield 8 Mapping IntrAgent RAGAvg. GPT-4o59/6560/6559.5/65 GPT-4.165/6561/6563/65 DeepSeek-R161/6561/6561/65 Table 5: Correctness agreement between human annota- tions and LLM-based mapping on the physics dataset with 65 instances. IntrAgent MappingGPT-4o GPT-4.1 DeepSeek-R1 RAG GPT-4o73.876.973.863.1 GPT-4.175.478.572.360.0 DeepSeek-R167.770.876.961.5 Table 6: Accuracy (%) ofIntrAgentfor different map- ping models compared with the Vanilla RAG baseline on the physics dataset. nearly identical accuracy, with GPT-4.1 achieving slightly better alignment (63 out of 65). Based on these results, we adopt GPT-4.1 as the default map- ping model for all main experiments. We hypothe- size that LLMs with stronger scientific reasoning and domain expertise could further improve map- ping qualityβan investigation we leave to future work. We also conduct a parallel study on the same physics dataset, where short-form answers are gen- erated byIntrAgentusing three different back- bone models, as well as by the vanilla RAG base- line. Each answer is then mapped to an MCQ option using the same prompt across three differ- ent mapping models. As shown in Table 6, al- though accuracy varies across backbone-mapping model combinations,IntrAgentconsistently out- performs the vanilla RAG baseline, which is the strongest among all baselines in our experiments. Robustness ofIntrAgentUnder Subpar Sec- tion Headings Here, we investigate an interest- ing edge case where a research paper fails to an- notate its sections with standard headings. We select a representative paper from the physics dataset (Rathnayake et al., 2024) and construct three alternative heading styles for the paper: a beginner-style rewrite, a highly noisy version, and a Shakespearean rendition. The full list of rewrites is shown in Appendix D.1. With GPT-4o as the backbone,IntrAgentachieves 89.2% accuracy on the original headings. For all three altered versions, it maintains strong performance, achieving 84.6% accuracy with only a modest drop. Appendix D.2 provides detailed reasoning traces illustrating how IntrAgenteffectively handles noisy headings dur- ing section ranking. These results address RQ3, demonstrating the robustness ofIntrAgentto variations in both eval- uation settings and input conditions. 7 Conclusion We introduceIntraView, a novel task that targets the critical yet underexplored practice of retrieving information from scientific literature. We introduce IntrAgent, the first LLM-based agent specifically designed forIntraView.IntrAgentmimics hu- man reading behavior by first identifying the most relevant sections and then iteratively extracting key information. To enable systematic evaluation of IntraViewapproaches, includingIntrAgent, we presentIntraBenchβ a benchmark spanning five high-impact scientific domains. Experimental re- sults show the superior performance ofIntrAgent over baselines on this benchmark. Limitations WhileIntrAgentandIntraBenchfocus on ad- vancing text-based scientific information retrieval and understanding, this work does not yet incorpo- rate non-textual modalities such as plots, figures, and tables. These visual elements often encapsu- late concentrated insights, including experimental trends, quantitative comparisons, and structural re- lationships, which can be essential for comprehen- sive scientific reasoning. In addition, the literature considered in this paper encompasses the major article types, but not all. For example, review papers are not included in our evaluation. In the future work, we will expand our benchmark to include more types of papers and queries. References Shubham Agarwal, Gaurav Sahu, Abhay Puri, Issam H Laradji, Krishnamurthy DJ Dvijotham, Jason Stan- ley, Laurent Charlin, and Christopher Pal. 2024. Litllm: A toolkit for scientific literature review. arXiv preprint arXiv:2402.01788. N. Anggriani, M. Z. Ndii, R. Amelia, and Sury. 2022. A mathematical covid-19 model considering asymp- tomatic and symptomatic classes with waning im- munity. Alexandria Engineering Journal, 61(1):113β 124. Epub 2021 May 14. 9 Shahzeb Ansari, Haiping Du, Fazel Naghdy, and David Stirling. 2022. Automatic driver cognitive fatigue detection based on upper body posture variations. Expert Systems with Applications, 203:117568. Anthropic. 2024.Introducing contextual retrieval. https://w.anthropic.com/engineering/ contextual-retrieval. Accessed: 2025-09-10. S. K. Biswas, J. K. Ghosh, S. Sarkar, and U. Ghosh. 2020. Covid-19 pandemic in india: a mathematical model study. Nonlinear Dynamics, 102(1):537β553. Epub 2020 Sep 21. Andres M Bran, Sam Cox, Oliver Schilter, Carlo Baldas- sari, Andrew D White, and Philippe Schwaller. 2024. Augmenting large language models with chemistry tools. Nature Machine Intelligence, 6(5):525β535. Tex.copyright: 2024 The Author(s). Jingyi Chai, Shuo Tang, Rui Ye, Yuwen Du, Xinyu Zhu, Mengcheng Zhou, Yanfeng Wang, Yuzhi Zhang, Lin- feng Zhang, Siheng Chen, and 1 others. 2025. Sci- master: Towards general-purpose scientific ai agents, part i. x-master as foundation: Can we lead on human- ityβs last exam? arXiv preprint arXiv:2507.05241. Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading wikipedia to answer open- domain questions. In Proceedings of the 55th an- nual meeting of the association for computational lin- guistics (volume 1: Long papers), pages 1870β1879, Vancouver, Canada. Association for Computational Linguistics. Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Wang. 2021. HybridQA: a dataset of multi-hop question answering over tabular and textual data. arXiv preprint. Pradeep Dasigi, Kyle Lo, Iz Beltagy, Arman Cohan, Noah A. Smith, and Matt Gardner. 2021. A dataset of information-seeking questions and answers anchored in research papers. arXiv preprint. J. T. Dillon. 1984. The classification of research ques- tions. Review of Educational Research, 54(3):327β 361. Song Ding, Lunhu Hu, Xing Pan, Dujun Zuo, and Li- uwang Sun. 2025. Assessing human situation aware- ness reliability considering fatigue and mood using eeg data: A bayesian neural network-bayesian net- work approach. Reliability Engineering & System Safety, 260:110962. M. Erzina, A. Trelin, O. Guselnikova, A. Skvortsova, K. Strnadova, V. Svorcik, and O. Lyutakov. 2022. Quantitative detection ofΞ±1-acid glycoprotein (agp) level in blood plasma using SERS and CNN transfer learning approach. Sensors and Actuators B: Chemi- cal, 368:132057. Received 25 Feb 2022, Revised 9 May 2022, Accepted 14 May 2022, Available online 20 May 2022, Version of Record 23 May 2022. Elena Escobar-Linero, Manuel DomΓnguez-Morales, and JosΓ© Luis Sevillano. 2022. Workerβs physical fatigue classification using neural networks. Expert Systems with Applications, 198:116784. Aniruddha Gaikwad, Reza Yavari, Mohammad Mon- tazeri, Kevin Cole, Linkan Bian, and Prahalada Rao. 2020. Toward the digital twin of additive manufac- turing: Integrating thermal simulations, sensing, and analytics to detect process faults. Iise Transactions, 52(11):1204β1217. Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. 2024. Retrieval-augmented gen- eration for large language models: a survey. arXiv preprint. Peng Gong, Wang , Jie, Yu , Le, Zhao , Yongchao, Zhao , Yuanyuan, Liang , Lu, Niu , Zhenguo, Huang , Xiaomeng, Fu , Haohuan, and Liu et al. , Shuang. 2013. Finer resolution observation and monitoring of global land cover: first mapping results with Landsat TM and ETM+ data. International Journal of Remote Sensing, 34(7):2607β2654. Peng Gong, Bin Chen, Xuecao Li, Han Liu, Jie Wang, Yuqi Bai, Jingming Chen, Xi Chen, Lei Fang, and Shuailong et al. Feng. 2020a.Mapping essen- tial urban land use categories in China (EULUC- China): preliminary results for 2018. Science Bul- letin, 65(3):182β187. Peng Gong, Xuecao Li, Jie Wang, Yuqi Bai, Bin Chen, Tengyun Hu, Xiaoping Liu, Bing Xu, Jun Yang, Wei Zhang, and Yuyu Zhou. 2020b. Annual maps of global artificial impervious area (GAIA) between 1985 and 2018. Remote Sensing of Environment, 236:111510. Lillemor R.-M. Hallberg. 2010. Some thoughts about the literature review in grounded theory studies. In- ternational Journal of Qualitative Studies on Health and Well-being, 5(3). Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng Thai, Junhao Shen, Jinyi Hu, Xu Han, Yu- jie Huang, and Yuxiang Zhang et al. 2024. Olympiad- Bench: a challenging benchmark for promoting AGI with olympiad-level bilingual multimodal scientific problems. arXiv preprint. Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, and Bing et al. Qin. 2025. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Sys- tems, 43(2):1β55. Enahoro A. Iboi, Oluwaseun Sharomi, Calistus N. Ngonghala, and Abba B. Gumel. 2020. Mathemati- cal modeling and analysis of COVID-19 pandemic in Nigeria. Mathematical biosciences and engineering: MBE, 17(6):7192β7220. 10 Ziyan Jiang, Xueguang Ma, and Wenhu Chen. 2024. Longrag: Enhancing retrieval-augmented generation with long-context llms. Preprint, arXiv:2406.15319. Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2021. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Ap- plied Sciences, 11(14):6421. Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Co- hen, and Xinghua Lu. 2019. PubMedQA: a dataset for biomedical research question answering. In Pro- ceedings of the 2019 conference on empirical meth- ods in natural language processing and the 9th inter- national joint conference on natural language pro- cessing (EMNLP-IJCNLP), pages 2567β2577, Hong Kong, China. Association for Computational Linguis- tics. Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. 2017. TriviaQA: a large scale distantly supervised challenge dataset for reading comprehen- sion. arXiv preprint. Muhammad Altaf Khan and Abdon Atangana. 2022. Mathematical modeling and analysis of COVID-19: A study of new variant Omicron. Physica A: Statisti- cal Mechanics and its Applications, 599:127452. Anastasia Krithara, Anastasios Nentidis, Konstantinos Bougiatiotis, and Georgios Paliouras. 2023. BioASQ- QA: A manually curated corpus for Biomedical Ques- tion Answering. Scientific Data, 10(1):170. Amit Kumar, Md Redwan Islam, Susu M. Zughaier, Xianyan Chen, and Yiping Zhao. 2024.Preci- sion classification and quantitative analysis of bac- teria biomarkers via surface-enhanced Raman spec- troscopy and machine learning. Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, 320:124627. Saeb Ragani Lamooki, Sahand Hajifar, Jiyeon Kang, Hongyue Sun, Fadel M Megahed, and Lora A Cavuoto. 2022. A data analytic end-to-end frame- work for the automated quantification of ergonomic risk factors across multiple tasks using a single wear- able sensor. Applied ergonomics, 102:103732. Yoonjoo Lee, Kyungjae Lee, Sunghyun Park, Dasol Hwang, Jaehyeon Kim, Hong-in Lee, and Moontae Lee. 2023. Qasa: advanced question answering on scientific articles. In International Conference on Machine Learning, pages 19036β19052. PMLR. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, and Vladimir et al. Karpukhin. 2020. Retrieval-augmented generation for knowledge- intensive NLP tasks. In Proceedings of the 34th international conference on neural information pro- cessing systems, Nips β20, Vancouver, BC, Canada and Red Hook, NY, USA. Curran Associates Inc. Jiarui Li, Ye Yuan, and Zehua Zhang. 2024. Enhancing LLM factual accuracy with RAG to counter halluci- nations: a case study on domain-specific queries in private knowledge-bases. arXiv preprint. Rui Li, Mingzhou Jin, and Vincent C Paquit. 2021a. Ge- ometrical defect detection for additive manufacturing with machine learning models. Materials & Design, 206:109726. Xiaoting Li, Tengyun Hu, Peng Gong, Shihong Du, Bin Chen, Xuecao Li, and Qi Dai. 2021b. Mapping essential urban land use categories in beijing with a fast area of interest (AOI)-based method. Remote Sensing, 13(3):477. Chin-Yew Lin. 2004. ROUGE: a package for auto- matic evaluation of summaries. In Text summariza- tion branches out, pages 74β81, Barcelona, Spain. Association for Computational Linguistics. Han Liu, Peng Gong, Jie Wang, Nicholas Clinton, Yuqi Bai, and Shunlin Liang. 2020. Annual dynam- ics of global land cover and its long-term changes from 1982 to 2015. Earth System Science Data, 12(2):1217β1243. Bo-Ru Lu, Nikita Haduong, Chien-Yu Lin, Hao Cheng, Noah A. Smith, and Mari Ostendorf. 2024a. Efficient encoder-decoder transformer decoding for decompos- able tasks. arXiv preprint. Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Fo- erster, Jeff Clune, and David Ha. 2024b. The AI scientist: Towards fully automated open-ended scien- tific discovery. arXiv preprint. Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chun- yuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. 2024c. MathVista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint. Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai- Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. Preprint, arXiv:2209.09513. Jakub LΓ‘la, Odhran OβDonoghue, Aleksandar Shtedrit- ski, Sam Cox, Samuel G. Rodriques, and Andrew D. White. 2023. PaperQA: Retrieval-augmented genera- tive agent for scientific research. arXiv preprint. Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct elec- tricity? A new dataset for open book question an- swering. arXiv preprint. George A. Miller. 1956. The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological Review, 63(2):81β97. 11 Niklas Muennighoff, SU Hongjin, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2024. Generative representational in- struction tuning. In The Thirteenth International Conference on Learning Representations. Aditya Nagori, Ricardo Accorsi Casonatto, Ayush Gau- tam, Abhinav Manikantha Sai Cheruvu, and Rishike- san Kamaleswaran. 2025. Open-source agentic hy- brid rag framework for scientific literature review. arXiv preprint arXiv:2508.05660. FaΓ―Γ§al NdaΓ―rou, IvΓ‘n Area, Juan J. Nieto, and Delfim F. M. Torres. 2020.Mathematical modeling of COVID-19 transmission dynamics with a case study of Wuhan. Chaos, Solitons & Fractals, 135:109846. Duc S. H. Nguyen, Bach G. Truong, Phuong T. Nguyen, Juri Di Rocco, and Davide Di Ruscio. 2025. Team- work makes the dream work: LLMs-based agents for GitHub readme.md summarization. Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic eval- uation of machine translation. In Proceedings of the 40th annual meeting of the association for com- putational linguistics, pages 311β318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. Xin Qi, Quyu Kong, Hongyi Ji, Yue Shen, Yuqi Liu, Yan Sun, Zhilin Zhang, Zhaorong Li, Xunlong Xia, and Bing et al. Deng. 2024. BioInformatics agent (BIA): Unleashing the power of large language models to reshape bioinformatics workflow. Tex.copyright: Β© 2024, Posted by Cold Spring Harbor Labora- tory. This pre-print is available under a Creative Commons License (Attribution-NonCommercial 4.0 International), C BY-NC 4.0, as described at http://creativecommons.org/licenses/by-nc/4.0/. Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you donβt know: Unanswerable questions for SQuAD. arXiv preprint. Rathnayake Rathnayake, Zhenghao Zhao, Nathan McLaughlin, Wei Li, Yan Yan, Liaohai Chen, Qian Xie, Christine Wu, Mathew Mathew, and Rong Wang. 2024. Machine learning enabled multiplex detec- tion of periodontal pathogens by surface-enhanced Raman spectroscopy. International Journal of Bio- logical Macromolecules, 257:128773. Siva Reddy, Danqi Chen, and Christopher D. Manning. 2019. CoQA: a conversational question answering challenge. arXiv preprint. David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Di- rani, Julian Michael, and Samuel R. Bowman. 2023. GPQA: a graduate-level google-proof Q&a bench- mark. arXiv preprint. Joshua C. Rothstein, Jiaheng Cui, Yanjun Yang, Xi- anyan Chen, and Yiping Zhao. 2024. Ultra-sensitive detection of PFASs using surface enhanced Raman scattering and machine learning: a promising ap- proach for environmental analysis. Sensors & Diag- nostics, 3(8):1272β1284. Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Zicheng Liu, and Emad Barsoum. 2025. Agent laboratory: Using llm agents as research assistants. Preprint, arXiv:2501.04227. Luis Javier Segura, Tianjiao Wang, Chi Zhou, and Hongyue Sun. 2021. Online droplet anomaly de- tection from streaming videos in inkjet printing. Ad- ditive Manufacturing, 38:101835. Seyyed Hadi Seifi, Wenmeng Tian, Haley Doude, Mark A Tschopp, and Linkan Bian. 2019. Layer- wise modeling and anomaly detection for laser-based additive manufacturing. Journal of Manufacturing Science and Engineering, 141(8):081013. Sneha Senapati, Manleen Kaur, Neetu Singh, Smita S. Kulkarni, and Jitendra Pratap Singh. 2024. Afford- able paper-based surface-enhanced raman scattering substrates containing silver nanorods using glancing- angle deposition for nosocomial infection detection. ACS Applied Nano Materials, 7(7):6736β6748. Yijia Shao, Yucheng Jiang, Theodore A. Kanell, Pe- ter Xu, Omar Khattab, and Monica S. Lam. 2024. Assisting in writing wikipedia-like articles from scratch with large language models.Preprint, arXiv:2402.14207. Chandan Kumar Singh, Devesh Kumar, Vipul Sanap, and Rajesh Sinha. 2025. Llm-rspf: Large language model-based robotic system planning framework for domain specific use-cases. In 2025 IEEE/CVF Win- ter Conference on Applications of Computer Vision (WACV), pages 7277β7286. Michael D Skarlinski, Sam Cox, Jon M Laurent, James D Braza, Michaela Hinks, Michael J Hammer- ling, Manvitha Ponnapati, Samuel G Rodriques, and Andrew D White. 2024. Language agents achieve superhuman synthesis of scientific knowledge. arXiv preprint arXiv:2409.13740. Bingyi Su, Liwei Qing, Lu Lu, SeHee Jung, Xiaolei Fang, and Xu Xu. 2024a. Enhancing data privacy in human factors studies with federated learning. Hu- man Factors, page 00187208251348025. Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. 2024b. DRAGIN: Dynamic Retrieval Augmented Generation based on the Information Needs of Large Language Models. arXiv preprint. Piwat Suppawittaya. 2021. The effectiveness of chunk- ing methods for enhancing short-term memory of textual information. Psychology and Education Jour- nal, 57:6313β6327. Peggy Tang, Junbin Gao, Lei Zhang, and Zhiyong Wang. 2023. Efficient and Interpretable Compressive Text Summarisation with Unsupervised Dual-Agent Rein- forcement Learning. arXiv preprint. 12 Xiangru Tang, Tianyu Hu, Muyang Ye, Yanjun Shao, Xunjian Yin, Siru Ouyang, Wangchunshu Zhou, Pan Lu, Zhuosheng Zhang, and Yilun Zhao et al. 2025. ChemAgent: Self-updating library in large language models improves chemical reasoning. arXiv preprint. William John Thrift and Regina Ragan. 2019. Quantifi- cation of analyte concentration in the single molecule regime using convolutional neural networks. ANA- LYTICAL CHEMISTRY, 91(21):13337β13342. Bin Wang, Chao Xu, Xiaomeng Zhao, Linke Ouyang, Fan Wu, Zhiyuan Zhao, Rui Xu, Kaiwen Liu, Yuan Qu, and Fukai Shang et al. 2024a. MinerU: An open- source solution for precise document content extrac- tion. Ke Wang, Houxing Ren, Aojun Zhou, Zimu Lu, Sichun Luo, Weikang Shi, Renrui Zhang, Linqi Song, Mingjie Zhan, and Hongsheng Li. 2023a. Math- Coder: Seamless code integration in LLMs for en- hanced mathematical reasoning. arXiv preprint. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023b. Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368. Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic com- pression of pre-trained transformers.Preprint, arXiv:2002.10957. Xiaohua Wang, Zhenghua Wang, Xuan Gao, Feiran Zhang, Yixin Wu, Zhibo Xu, Tianyuan Shi, Zhengyuan Wang, Shizheng Li, Qi Qian, Ruicheng Yin, Changze Lv, Xiaoqing Zheng, and Xuan- jing Huang. 2024b.Searching for best prac- tices in retrieval-augmented generation. Preprint, arXiv:2407.01219. Zheng Wang, Zhongyang Li, Zeren Jiang, Dandan Tu, and Wei Shi. 2024c. Crafting personalized agents through retrieval-augmented generation on editable memory graphs. arXiv preprint. Zheng Wang, Shu Teo, Jieer Ouyang, Yongjun Xu, and Wei Shi. 2024d. M-RAG: Reinforcing large language model performance through retrieval-augmented gen- eration with multiple partitions. In Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers), pages 1966β1978, Bangkok, Thailand. Association for Computational Linguistics. Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing multiple choice science questions. arXiv preprint. Sharon Whitfield, , and Melissa A. Hofmann. 2023. Elicit: AI literature review research assistant. Public Services Quarterly, 19(3):201β207. Dayong Wu, Jiaqi Li, Baoxin Wang, Honghong Zhao, Siyuan Xue, Yanjie Yang, Zhijun Chang, Rui Zhang, Li Qian, and Bo Wang et al. 2024a. SparkRA: a retrieval-augmented knowledge service system based on spark large language model. In Proceedings of the 2024 conference on empirical methods in nat- ural language processing: System demonstrations, pages 382β389, Miami, Florida, USA. Association for Computational Linguistics. Shangyu Wu, Ying Xiong, Yufei Cui, Haolun Wu, Can Chen, Ye Yuan, Lianming Huang, Xue Liu, Tei-Wei Kuo, and Nan Guan et al. 2025. Retrieval-augmented generation for natural language processing: a survey. arXiv preprint. Yiran Wu, Feiran Jia, Shaokun Zhang, Hangyu Li, Erkang Zhu, Yue Wang, Yin Tat Lee, Richard Peng, Qingyun Wu, and Chi Wang. 2024b. MathChat: Con- verse to tackle challenging math problems with LLM agents. arXiv preprint. Sarjana Yadav, Sneha Senapati, Samir Kumar, Shashank K. Gahlaut, and Jitendra P. Singh. 2022. GLAD based advanced nanostructures for diversified biosensing applications: Recent progress. Biosen- sors, 12(12):1115. Yanjun Yang, Jiaheng Cui, Dan Luo, Jackelyn Mur- ray, Xianyan Chen, Sebastian HΓΌlck, Ralph A. Tripp, and Yiping Zhao. Rapid detection of SARS-CoV- 2 variants using an angiotensin-converting enzyme 2-based surface-enhanced raman spectroscopy sen- sor enhanced by CoVari deep learning algorithms. 9(6):3158β3169. Publisher: American Chemical So- ciety. Yanjun Yang, Hao Li, Les Jones, Jackelyn Murray, James Haverstick, Hemant K. Naikare, Yung-Yi C. Mosley, Ralph A. Tripp, Bin Ai, and Yiping Zhao. 2023. Rapid detection of SARS-CoV-2 RNA in human nasopharyngeal specimens using surface- enhanced raman spectroscopy and deep learning al- gorithms. ACS Sensors, 8(1):297β307. Fuda Ye, Shuangyin Li, Yongqi Zhang, and Lei Chen. 2024. R^2AG: Incorporating retrieval information into retrieval augmented generation. arXiv preprint. Antonio Jimeno Yepes, Yao You, Jan Milczek, Sebas- tian Laverde, and Renyu Li. 2024. Financial report chunking for effective retrieval augmented genera- tion. Preprint, arXiv:2402.05131. Da Yin, Faeze Brahman, Abhilasha Ravichander, Khy- athi Chandu, Kai-Wei Chang, Yejin Choi, and Bill Yuchen Lin. 2024. Agent Lumos: Unified and Modular Training for Open-Source Language Agents. arXiv preprint. Haining Zhang, Jingyuan Huang, Xiaoge Zhang, and Chak-Nam Wong. 2024a. Autonomous optimization of process parameters and in-situ anomaly detection in aerosol jet printing by an integrated machine learn- ing approach. Additive Manufacturing, 86:104208. 13 Huan Zhang, Yu Song, Ziyu Hou, Santiago Miret, and Bang Liu. 2024b. HoneyComb: a flexible LLM- based agent system for materials science. arXiv preprint. Yu Zhang, Xiusi Chen, Bowen Jin, Sheng Wang, Shui- wang Ji, Wei Wang, and Jiawei Han. 2024c. A comprehensive survey of scientific large language models and their applications in scientific discovery. In Proceedings of the 2024 conference on empiri- cal methods in natural language processing, pages 8783β8817, Miami, Florida, USA. Association for Computational Linguistics. Yiping Zhao, Amit Kumar, and Yanjun Yang. 2024. Unveiling practical considerations for reliable and standardized SERS measurements: lessons from a comprehensive review of oblique angle deposition- fabricated silver nanorod array substrates. Chemical Society Reviews, 53(2):1004β1057. Zhiying Zhu, Yiming Yang, and Zhiqing Sun. 2024. HaluEval-wild: Evaluating hallucinations of lan- guage models in the wild. arXiv preprint. 14 A Details about IntraBench The benchmark was created through a systematic process: first selecting representative science litera- ture, then manually crafting questions that span dif- ferent scientific four task-oriented categories: study subject & experimental setup, data characteristics & collection, technical approach & details, con- clusions & results. Special attention was given to ensuring professional quality and coverage across multiple subfields. A.1 Research Fields in IntraBench Surface-enhanced Raman Spectroscopy (SERS) Surface-enhanced raman spectroscopy(SERS) rep- resents a significant advancement in applied physics, particularly within the domain of optical physics. At its core, SERS is an extension of Ra- man spectroscopy, a technique that analyzes the inelastic scattering of light to provide insights into molecular vibrations. The enhancement in SERS arises when molecules are adsorbed onto nanos- tructured metallic surfaces, such as gold or silver, leading to a substantial increase in the Raman sig- nal. This amplification is primarily attributed to localized surface plasmon resonancesβcoherent oscillations of conduction electrons at the metal surface excited by incident lightβwhich gener- ate intense electromagnetic fields near the surface, thereby boosting the Raman scattering efficiency of nearby molecules. The remarkable sensitivity of SERS enables the detection of analytes at extremely low concentrations, down to the single-molecule level. This capability is particularly valuable in various engineering applications where identify- ing trace amounts of substances is crucial. For in- stance, in public safety, SERS-based sensors have been developed for the rapid and accurate detec- tion of bio-hazardous materials, including chemi- cal agents, viruses, and bacteria(Thrift and Ragan, 2019; Rathnayake et al., 2024; Yang et al., 2023; Erzina et al., 2022). Such applications benefit from SERSβs ability to provide specific molecular finger- prints, facilitating the identification of dangerous substances even in complex environments. Remote SensingLand cover and land use classi- fication is a central research direction in remote sensing science, aiming to automatically iden- tify and spatially represent surface types and hu- man activity patterns based on multi-source re- mote sensing data(Gong et al., 2020b; Liu et al., 2020; Gong et al., 2013; Li et al., 2021b; Gong et al., 2020a). This field widely utilizes medium- and high-resolution satellite imagery from optical sensors (e.g., Landsat, Sentinel-2) and synthetic aperture radar (e.g., Sentinel-1), employing tech- niques such as supervised classification, object- based analysis, time-series processing, and deep learning for high-accuracy mapping. In recent years, deep convolutional neural networks such as U-Net, DeepLab, and HRNet have been exten- sively applied to semantic segmentation of high- resolution imagery, significantly improving the ex- traction accuracy of fine-scale urban features such as buildings, roads, and impervious surfaces. In addition, the integration of optical and radar data with auxiliary variablesβsuch as nighttime lights and population densityβhas enhanced model per- formance in complex urban environments and facil- itated dynamic monitoring of land use changes. These studies are of great value for supporting global change research, environmental assessment, and sustainable urban planning. Infectious-disease ModelingInfectious-disease modeling and forecasting play a central role in public health by enabling researchers and poli- cymakers to anticipate the spread and burden of infectious diseases. Among the tools available, mathematical modelingβparticularly compartmen- tal models such as SIR and SEIRβhas proven in- dispensable for simulating disease transmission dy- namics, assessing intervention strategies, and guid- ing resource allocation(Khan and Atangana, 2022; Biswas et al., 2020; Anggriani et al., 2022; Iboi et al., 2020; NdaΓ―rou et al., 2020). The COVID- 19 pandemic has exemplified the value of these models, as they have been critical for estimating infection trajectories, evaluating the timing and effectiveness of control measures, and informing real-time policy responses during a rapidly evolv- ing global crisis. Accurate modeling has thus been essential for understanding and managing the im- pact of COVID-19 across different settings and timeframes. Human Factor Human factors is a cornerstone of industrial engineering, focusing on the design and evaluation of systems that balance human well- being and overall system performance. Human factors has addressed challenges in physical work- load, cognitive workload, and humanβmachine in- teraction across domains such as manufacturing, transportation, and healthcare. A critical area is the classification and prediction of human fatigue and 15 risk. This is because prolonged exposure to repeti- tive motions, awkward postures, or high-intensity workloads can impair human performance, increase error rates, and elevate injury risk. Recent studies have leveraged wearable sensors, biomechanical modeling, and machine learning methods to detect early indicators of fatigue and quantify risks, en- abling proactive intervention(Lamooki et al., 2022; Escobar-Linero et al., 2022; Ansari et al., 2022; Ding et al., 2025; Su et al., 2024a). Such ap- proaches underscore the evolving role of human factors research in advancing resilient, human- centered industrial systems. Additive Manufacturing Additive manufactur- ing (AM), rooted in materials science and manufac- turing engineering, refers to a family of processes that build components layer by layer directly from digital models. Over the past decades, AM has evolved from rapid prototyping into a transforma- tive production technology, enabling complex ge- ometries, lightweight structures, and customized products across aerospace, biomedical, and energy sectors. While the promise of AM lies in its design freedom and material efficiency, ensuring consis- tent quality remains a central challenge. Quality control in AM encompasses detecting defects such as porosity and dimensional inaccuracy, which can critically impact the structural integrity and perfor- mance of printed parts. Researchers increasingly employ in-situ monitoring, computer vision, and machine learning-based defect detection to address variability AM processes(Zhang et al., 2024a; Li et al., 2021a; Seifi et al., 2019; Segura et al., 2021; Gaikwad et al., 2020). Such quality-control ef- forts are pivotal to advancing AM toward reliable, industrial-scale deployment. A.2 Research Literature in IntraBench Table 7 contains all 25 papers, five per field, used in IntraBench. A.3 Research Questions in IntraBench In total,IntraBenchcomprises 63 research ques- tions, systematically categorized into four task- oriented categories: (1) Study subject & experi- mental setup, (2) Data characteristics & collection, (3) Technical approach & details, and (4) Conclu- sions & results. These questions are designed to elicit critical information from scientific papers across five domains. Specifically, the SERS in chemistry physics dataset contributes 13 questions, the infectious-disease modeling in public health dataset contributes 11 questions, the remote sens- ing in earth science dataset contributes 12 ques- tions, the human performance sensing in engineer- ing dataset contributes 13 questions, and the ad- ditive manufacturing in material science dataset contributes 14 questions. The complete lists are provided in Table 8, Table 9, Table 10, Table 11, and Table 12, respectively. Across all domains, the 63 questions remain balanced across the first three task-oriented cat- egoriesβStudy subject & experimental setup, Data characteristics & collection, and Technical ap- proach & detailsβeach covering foundational as- pects of experimental design, data acquisition, and analytical methodology. The remaining questions belong to the Conclusions & results category, em- phasizing the outcome evaluation and performance reporting aspects of scientific research. While the overall structure 63 consistent, each domain reflects distinct emphases. The infectious- disease modeling dataset prioritizes experimental setup due to the importance of compartmental struc- tures and intervention parameters. The remote sens- ing dataset highlights data characteristics such as sensor type, spatial resolution, and temporal cover- age. The SERS in physics dataset shows a balanced distribution aligned with typical SERS workflows from substrate preparation to data-driven analysis. The humanfactor in engineering dataset empha- sizes human performance measurement and mul- timodal sensing setups, reflecting the diversity of sensor placements and physiological variables in human studies. The Additive manufacturing in material science dataset focuses on process moni- toring and machine learning evaluation within addi- tive manufacturing workflows, emphasizing mate- rialβprocessβdefect relationships and reproducible modeling strategies. To ensure usability and reproducibility, each re- search question was expert-annotated with detailed explanatory notes. A.4 Main Result Report over Task-oriented Categories To further analyze performance differences among backbone large language models (LLMs), we eval- uateIntrAgentacross four task-oriented cate- gories: (1) Study subject & experimental setup (S&E), (2) Data characteristics & collection (D&C), (3) Technical approach & details (T&D), and (4) Conclusions & results (C&R). 16 TitleRef. Public Health - Infectious-disease Modeling Mathematical modeling and analysis of COVID-19: a study of new variant Omicron(Khan and Atangana, 2022) COVID-19 pandemic in India: a mathematical model study(Biswas et al., 2020) A mathematical COVID-19 model considering asymptomatic and symptomatic classes with waning immunity (Anggriani et al., 2022) Mathematical modeling and analysis of COVID-19 pandemic in Nigeria(Iboi et al., 2020) Mathematical modeling of COVID-19 transmission dynamics with a case study of Wuhan(NdaΓ―rou et al., 2020) Physics - Surface Enhanced Raman Spectroscopy Quantification of analyte concentration in the single molecule regime using convolutional neural networks (Thrift and Ragan, 2019) Machine learning enabled multiplex detection of periodontal pathogens by surface- enhanced Raman spectroscopy (Rathnayake et al., 2024) Rapid detection of SARS-CoV-2 variants using an angiotensin-converting enzyme 2- based surface-enhanced Raman spectroscopy sensor enhanced by CoVari deep learning algorithms (Yang et al.) Rapid detection of SARS-CoV-2 RNA in human nasopharyngeal specimens using surface-enhanced Raman spectroscopy and deep learning algorithms (Yang et al., 2023) Quantitative detection ofΞ±1-acid glycoprotein (AGP) level in blood plasma using SERS and CNN transfer learning approach (Erzina et al., 2022) Earth Science - Remote Sensing Annual maps of global artificial impervious area (GAIA) between 1985 and 2018(Gong et al., 2020b) Annual dynamics of global land cover and its long-term changes from 1982 to 2015(Liu et al., 2020) Finer resolution observation and monitoring of global land cover: first mapping results with Landsat TM and ETM+ data (Gong et al., 2013) Mapping essential urban land use categories in Beijing with a fast area of interest (AOI)- based method (Li et al., 2021b) Mapping essential urban land use categories in China (EULUC-China): preliminary results for 2018 (Gong et al., 2020a) Engineering - Human Factor A data analytic end-to-end framework for the automated quantification of ergonomic risk factors across multiple tasks using a single wearable sensor (Lamooki et al., 2022) Assessing human situation awareness reliability considering fatigue and mood using EEG data: a Bayesian neural network-Bayesian network approach (Ding et al., 2025) Automatic driver cognitive fatigue detection based on upper body posture variations(Ansari et al., 2022) Enhancing data privacy in human factors studies with federated learning(Su et al., 2024a) Workerβs physical fatigue classification using neural networks(Escobar-Linero et al., 2022) Material Science - Additive Manufacturing Autonomous optimization of process parameters and in-situ anomaly detection in aerosol jet printing by an integrated machine learning approach (Zhang et al., 2024a) Geometrical defect detection for additive manufacturing with machine learning models(Li et al., 2021a) Layer-wise modeling and anomaly detection for laser-based additive manufacturing(Seifi et al., 2019) Online droplet anomaly detection from streaming videos in inkjet printing(Segura et al., 2021) Toward the digital twin of additive manufacturing: integrating thermal simulations, sensing, and analytics to detect process faults (Gaikwad et al., 2020) Table 7: Literature across five scientific domains used in our benchmark. 17 Task-orientedCate- gory Research Question Q Study subject & experimental setup What are the main analytes type studied? What are the material and structure, or morphology of the SERS substrates used? How many analytes are investigated? Data characteristics & collection What is the excitation laser wavelength used for SERS measurements? What is the spectral range collected for the analysis of the analytes? How many spectra are collected per analyte under each experimental condition? How many experimental replications are conducted to ensure reproducibility? Technical approach & details What is the primary machine learning task addressed in this study? Which machine learning algorithm is implemented? What data splitting strategy is applied, and the parameters? How many epochs are used during model training? Conclusions & results What performance metrics are employed to evaluate the machine learning mod- els? What are the reported performance values? Table 8: Research Questions in IntraBench: SERS in chemistry physics (Phys) Task-orientedCate- gory Research Question Q Study subject & experimental setup Into how many compartments is the population divided in the model? What is the initial susceptible population of the model? What is the initial infected population of the model? How many interventions are addressed in the paper? Data characteristics & collection What is the source location or country of origin for the data used in this study? Technical approach & details What is the model used in this paper? What is the transmission rate? What is the disease-induced mortality rate? What values of the basic reproduction number were considered in the model? Conclusions & results What are the novel contributions of the paper? What are the limitations of the paper? Table 9: Research Questions in IntraBench: infectious-disease modeling in Public Health (PH) These evaluations reveal how each backbone contributes to different aspects of literature rea- soning. Figure 3 visualizes the results across seven LLMs. Two overall trends emerge. First, IntrAgentmaintains balanced performance across all categories regardless of the backbone, con- firming its robustness and model-agnostic design. Second, although individual models show slight strengths in specific areas, the overall variation re- mains moderate, indicating that the contextual read- ing and iterative synthesis strategy ofIntrAgent mitigates backbone dependency. In detail, GPT-4.1 achieves the highest overall balance (S&E = 66.8, D&C = 72.7, T&D = 83.4, C&R = 64.3), especially excelling in technical rea- soning. o4-mini performs comparably (65.7, 83.3, 77.5, 72.3), outperforming GPT-4.1 in D&C and C&R. Gemini 2.5 Pro maintains uniformly high results (76.7, 76.0, 75.0, 64.8), reflecting strong generalization. DeepSeek-R1 (56.4, 71.8, 82.6, 67.9) and o3 (60.0, 76.7, 82.5, 68.0) show stable interpretative abilities, while Llama-3.1-70B (62.2, 65.6, 77.9, 75.0) exhibits strength in conclusion- oriented reasoning. Overall, GPT-4.1 and o4-mini demonstrate the most consistent and well-rounded performance, while Gemini 2.5 Pro and Llama-3.1-70B show domain-specific advantages. These results suggest thatIntrAgentβs architecture effectively harmo- nizes reasoning depth and retrieval accuracy across heterogeneous LLMs, ensuring adaptability with- out overfitting to any specific model. B Additional Experiments and Reports B.1 Main Experiment - Domain-level Accuracy Breakdown Table 2 in the main text reports averaged accura- cies for every baselineβLLM combinations. For 18 Task-orientedCate- gory Research Question Q Study subject & experimental setup What is the number of land-cover / land-use classes classified in this study? What is the spatial extent of the study area? What is the geographic type of the study area? Data characteristics & collection What is the temporal scope of the data used? What type of remote sensing data is used? Which specific satellite data is used? What is the spatial resolution of the primary imagery used? Are auxiliary features used beyond raw spectral bands? Technical approach & details What type of model is implemented in this study? What performance metrics are reported? Conclusions & results Is any comparative analysis included? What is the reported overall accuracy (OA)? Table 10: Research Questions in IntraBench: remote sensing in Earth Science (ES) Task-orientedCate- gory Research Question Q Study subject & experimental setup What are the subjectsβ occupational roles? What specific task or activity are the subjects performing? What is the study context or environment in which participants perform the tasks? If the data are referenced from prior work, please indicate the source. Data characteristics & collection What are the primary sensing modalities or measurement instruments employed in the study to capture human performance and physiological responses? What is the anatomical or body placement of the sensors used in the study? What is the sampling rate (Hz)? What is the total number of participants involved in the study? Technical approach & details How are the participantsβ physical, cognitive, or perceptual states assessed or reflected in the study? What is the primary modeling objective in this study? What is the data partitioning strategy used during model training, and what are the parameters? What is the number of epochs used during model training (i.e., how many complete passes through the entire training dataset)? Conclusions & results Which performance metrics are used to assess the effectiveness of the machine learning models? What are the reported values for the performance metrics used to evaluate the machine learning models? Table 11: Research Questions in IntraBench: Human performance sensing in engineering (Engr) each pair we compute the simple arithmetic mean over the five domain datasets shown below, so ev- ery domain carries equal weight: SERS in physics, infectious-disease modeling in public health, re- mote sensing in earth science, human performance sensing in engineering, additive manufacturing in material science. While the main table focuses on these ag- gregated averages, the following tables (Ta- ble 13, 14, 15, 16, 17) present the detailed per- domain accuracies for all baselineβLLM combina- tions. Each table corresponds to one of the five domains listed above. Boldface highlights the best score within each domain, allowing readers to ex- amine where specific retrieval strategies or LLM backbones perform most effectively. B.2 Statistical Analysis B.2.1 Significance Testing of IntrAgent Performance To rigorously assess whetherIntrAgentachieves statistically significant performance gains over existing baselines, we conducted a one-sided Wilcoxon signed-rank test across all five domains and seven backbones (n = 35paired samples per baseline). For each baseline, accuracies were paired with those ofIntrAgenton identical back- bones within the same domain. 19 Task-orientedCate- gory Research Question Q Study subject & experimental setup What type of additive manufacturing process is studied? What type of material is used for printing? What kind of shape or product is printed? What is the primary defect being studied? Data characteristics & collection What sensors are used to measure the process? What is the sampling rate (Hz)? If relevant, what is the spatial resolution (ΞΌm)? Technical approach & details What is the machine learning objective in this study? What machine learning algorithm is used? How are the data split during machine learning? How many replications are conducted during machine learning, if any? How many epochs are used during machine learning, if any? Conclusions & results What metrics are used to evaluate the machine learning models? What are the values for these metrics? Table 12: Research Questions in IntraBench: Additive manufacturing in material science (MS) BaselineGPT-4o GPT-4.1 DS-R1o3o4-mini Gemini-2.5 Pro Llama-3.1-70B RAGVanilla RAG all-MiniLM-L6-v260.066.264.658.560.060.067.7 Vanilla RAG E5-mistral-7b-instruct61.570.763.863.164.664.666.2 Vanilla RAG GritLM-7B60.067.658.558.560.058.564.6 Contextual RAG E5-mistral-7b-instruct58.566.258.558.560.053.958.5 Contextual RAG GritLM-7B53.967.664.658.560.060.058.5 DRAGIN38.543.047.743.146.244.646.2 R 2 AG58.556.958.556.956.956.955.4 LongRAG64.670.764.656.958.555.456.9 AgentLUMOS52.350.850.853.856.956.956.9 PaperQA249.250.852.349.252.350.852.3 Agentic-Hybrid-RAG60.063.161.555.455.455.453.8 SciMaster58.558.563.153.858.555.455.4 IntrAgent (Ours)75.478.578.576.981.578.572.3 Table 13: Accuracies (%) on IntraBench, Physics BaselineGPT-4o GPT-4.1 DS-R1o3o4-mini Gemini-2.5 Pro Llama-3.1-70B RAGVanilla RAG all-MiniLM-L6-v252.760.065.658.261.858.254.5 Vanilla RAG E5-mistral-7b-instruct52.761.765.654.656.454.658.2 Vanilla RAG GritLM-7B52.761.758.256.452.754.558.2 Contextual RAG E5-mistral-7b-instruct56.458.365.658.254.556.456.4 Contextual RAG GritLM-7B58.358.354.656.458.256.454.5 DRAGIN41.840.047.341.847.343.643.6 R 2 AG54.558.261.850.949.150.950.9 LongRAG50.954.660.047.349.150.947.3 AgentLUMOS43.647.349.149.147.347.352.7 PaperQA238.245.547.347.345.547.350.9 Agentic-Hybrid-RAG54.558.260.052.752.750.950.9 SciMaster58.254.560.050.950.949.150.9 IntrAgent (Ours)58.269.169.165.567.370.969.1 Table 14: Accuracies (%) on IntraBench, Public Health Hypotheses H 0 : Median(IntrAgentβ Baseline) = 0, H 1 : Median(IntrAgentβ Baseline) > 0. HereH 0 indicates no improvement ofIntrAgent over the baseline, whileH 1 tests for significantly higher performance, result shown 18. Interpretation Across all baselines, the me- dian performance improvements range from +10.9to+27.3percentage points. The Holmβ Bonferroniβadjustedp-values (p adj < 10 β6 ) in- dicate that these gains are highly significant and not attributable to random variation. Even after correction for multiple comparisons, all12tests reject the null hypothesisH 0 atΞ± = 0.05. These 20 BaselineGPT-4o GPT-4.1 DS-R1o3o4-mini Gemini-2.5 Pro Llama-3.1-70B RAGVanilla RAG all-MiniLM-L6-v260.058.360.058.361.755.060.0 Vanilla RAG E5-mistral-7b-instruct60.060.058.358.361.751.760.0 Vanilla RAG GritLM-7B61.760.058.353.351.751.761.7 Contextual RAG E5-mistral-7b-instruct58.361.860.051.751.758.361.7 Contextual RAG GritLM-7B58.358.260.048.356.758.360.0 DRAGIN40.043.340.043.345.046.746.7 R 2 AG58.356.758.355.055.053.360.0 LongRAG60.063.363.353.353.353.358.3 AgentLUMOS46.751.746.753.356.751.755.0 PaperQA251.750.053.350.045.048.355.0 Agentic-Hybrid-RAG58.358.358.356.755.055.060.0 SciMaster56.756.760.056.758.360.061.7 IntrAgent (Ours)63.370.070.070.069.170.165.0 Table 15: Accuracies (%) on IntraBench, Earth Science BaselineGPT-4o GPT-4.1 DS-R1o3o4-mini Gemini-2.5 Pro Llama-3.1-70B RAGVanilla RAG all-MiniLM-L6-v264.661.560.058.555.461.552.3 Vanilla RAG E5-mistral-7b-instruct60.058.561.560.060.060.055.4 Vanilla RAG GritLM-7B63.156.961.561.563.163.155.4 Contextual RAG E5-mistral-7b-instruct63.164.360.061.556.963.152.3 Contextual RAG GritLM-7B66.264.360.060.058.463.160.0 DRAGIN47.750.852.346.244.644.646.2 R 2 AG60.061.563.160.056.956.956.9 LongRAG67.769.270.863.166.263.161.5 AgentLUMOS55.458.564.658.556.958.553.8 PaperQA250.852.358.553.847.752.352.3 Agentic-Hybrid-RAG64.661.566.261.563.161.558.5 SciMaster64.660.066.261.561.560.056.9 IntrAgent (Ours)80.081.578.581.575.476.966.2 Table 16: Accuracies (%) on IntraBench, Engineering BaselineGPT-4o GPT-4.1 DS-R1o3o4-mini Gemini-2.5 Pro Llama-3.1-70B RAGVanilla RAG all-MiniLM-L6-v264.360.071.468.668.674.361.4 Vanilla RAG E5-mistral-7b-instruct62.970.070.065.764.368.662.9 Vanilla RAG GritLM-7B64.370.071.468.664.364.367.1 Contextual RAG E5-mistral-7b-instruct67.168.670.065.768.662.965.7 Contextual RAG GritLM-7B67.165.768.668.670.070.062.9 DRAGIN44.345.747.145.751.450.044.3 R 2 AG65.764.365.760.058.660.057.1 LongRAG67.165.768.664.364.362.962.9 AgentLUMOS52.957.165.761.464.360.061.4 PaperQA248.645.758.658.655.757.158.6 Agentic-Hybrid-RAG61.460.065.761.462.961.460.0 SciMaster57.158.667.162.961.461.460.0 IntrAgent (Ours)72.980.075.772.975.782.971.4 Table 17: Accuracies (%) on IntraBench, Material Science findings confirm thatIntrAgentconsistently and significantly outperforms every baseline across do- mains and backbones, demonstrating robust, model- agnostic performance advantages. B.2.2Statistical Analysis of Paper Length and IntrAgent Performance IntraBench spans a wide range of paper lengths, with an average of 15.2 pages (SD = 9.04), a me- dian of 13 pages, a minimum of 6 pages, and a maximum of 49 pages. To examine whether pa- per length affects agent performance, we build a mixed-effects linear regression model with the per- paper accuracy ofIntrAgentas the response and paper length as the predictor. Since each paper is evaluated using three backbone LLMs (GPT-4o, GPT-4.1, and DeepSeek-R1), we treat backbone choice as repeated measurements. 21 BaselineMedian β(%) p raw p adj Vanilla RAG all-MiniLM-L6-v2+11.72.9Γ 10 β11 3.5Γ 10 β10 Vanilla RAG E5-mistral-7b-instruct+10.92.9Γ 10 β11 3.5Γ 10 β10 Vanilla RAG GritLM-7B+11.72.9Γ 10 β11 3.5Γ 10 β10 Contextual RAG E5-mistral-7b-instruct+12.82.9Γ 10 β11 3.5Γ 10 β10 Contextual RAG GritLM-7B+12.95.8Γ 10 β11 3.5Γ 10 β10 DRAGIN+27.32.9Γ 10 β11 3.5Γ 10 β10 R 2 AG+16.82.9Γ 10 β11 3.5Γ 10 β10 LongRAG+12.32.9Γ 10 β11 3.5Γ 10 β10 LUMOS+18.52.9Γ 10 β11 3.5Γ 10 β10 PaperQA2+21.82.9Γ 10 β11 3.5Γ 10 β10 Agentic-Hybrid-RAG+14.12.9Γ 10 β11 3.5Γ 10 β10 SciMaster+14.61.8Γ 10 β7 1.8Γ 10 β7 Table 18: Wilcoxon signed-rank test comparingIntrAgentwith all baselines across 35 paired samples (5 domains Γ7 backbones). All tests are one-sided (H 1 :IntrAgent >baseline). Reportedp-values are HolmβBonferroni corrected for multiple comparisons. All comparisons remain significant at p adj < 0.001. Figure 3: Radar plot comparing the performance of seven backbone LLMsβGPT-4o, GPT-4.1, DeepSeek- R1, o3, o4-mini, Gemini 2.5 Pro, and Llama-3.1- 70Bβacross four research-question categories: Study subject & experimental setup (S&E), Data character- istics & collection (D&C), Technical approach & de- tails (T&D), and Conclusions & results (C&R). GPT-4.1 and o4-mini show the strongest overall balance, while Gemini 2.5 Pro and Llama-3.1-70B demonstrate stable domain-specific strengths. The plot highlights that de- spite moderate variation across backbones,IntrAgent remains consistently robust across all task categories. Hypotheses H 0 : Ξ² length = 0, H 1 : Ξ² length ΜΈ= 0. Here,Ξ² length denotes the regression coefficient as- sociated with paper length.H 0 indicates that paper length has no effect onIntrAgentβs performance, whileH 1 tests whether paper length has a statisti- cally significant effect. Interpretation The model estimates a slope co- efficient ofβ0.1444for paper length, withp = 0.463 > 0.05 and a 95% confidence interval of (β0.535, 0.246). Although the estimated coeffi- cient is negative, its magnitude is small, and thep- value is well above the standard significance thresh- old ofΞ± = 0.05. Therefore, we fail to rejectH 0 , indicating that paper length is not significantly as- sociated withIntrAgentβs performance in the cur- rent evaluation. B.3 Execution Time Report In this section, we report the computational runtime required to complete a single test instance on av- erage. The runtime ofIntrAgentis influenced by several factors, including the number of reasoning iterations until information sufficiency is achieved, API connection stability, and the response speed of the backbone LLM. Nevertheless, we provide an approximate runtime in seconds, averaged over multiple runs. (See Table19) Since the number of runs can vary across exper- iments, we place greater emphasis on the median runtime on the default setting to provide a more rep- resentative estimate.IntrAgentachieves a median runtime of approximately 129 seconds, or about two minutes per paper per question. In comparison, the baseline methods require slightly less than one minute. Given the nature of theIntraViewtask, our pri- mary concern is accuracy rather than speed. Never- theless, the agent remains significantly faster than human-level reading and reasoning over multiple full-text papers. 22 MethodStepTime estimation (sec) Evaluation protocolMCQ mapping5 IntrAgentFormat conversion36 Section Ranking stage4 Iterative Reading stage (1 iter)42 Median (Default setting)129 Average (Default setting)259 Vanilla RAGPlaintext conversion2 Short answer generation35 DRAGINPlaintext conversion2 Paragraph-Level Corpus Construction4 Chain-of-Thought Construction(each one)23 Step-by-step Reasoning55 R 2 AGPlaintext conversion2 Retriever Embedding2 Ranking5 Reranking10 Answer Generation with Retrieval Guidance30 LumosPlaintext conversion2 Planning for a single Subgoal6 Converting Subgoal into a Retrieval Action2 Execution2 Table 19: Estimated runtime (in seconds) forIntrAgentand baselines. Running with GPT-4o. Values are averaged over multiple runs per test instance. C Ablation Studies Detail C.1 Failure Case Study C.1.1 Hallucination Failure Case from Missing Information Sufficiency Check As an extension of our ablation study, we examine a specific failure mode caused by removing the information sufficiency check from the Iterative Reading stage. This case highlights the critical role of sufficiency verification in preventing answer hallucination. We analyze a representative question from the SERS dataset based on the paper by Rathnayake et al. (Rathnayake et al., 2024): The correct section is successfully retrieved by section ranking stag: Ex- perimental - Surface entrapping of AGP and SERS measurements. The following detail is extracted: βThe measurements were performed on a ProRaman-L spectrometer with 785 nm excitation wavelength using a40Γ objective, under the following condi- tions: power density of approximately 40ΞΌW/ΞΌm 2 , accumulation time of40 s, averaging over 3 spectra, and a spectral range of 400β3000 cm β1 .β This evidence directly supports the correct an- swer: 785 nm. In contrast, when the information sufficiency check is disabled, the agent outputs a fully halluci- nated answer: βThe excitation laser wavelength used for SERS measurements is confirmed to be 532 nm. This is supported by the direct statement: βThe excitation wavelength for the SERS measurements is 532 nmβ found in the text (Chunk 3, Section_title: Methodology).β In this case, the cited sentence, section title, and numerical value are entirely fabricated and do not exist in the original document. The agent not only provides an incorrect answer but does so with un- justified confidence and fabricated attribution. This failure mode demonstrates how critical the information sufficiency check is for ensuring fac- tual grounding. Its absence increases the risk of hal- lucination, where the system generates plausible- sounding yet entirely unsupported responses. C.1.2Contextual Relationship Blindness from Domain-Specific Knowledge Gaps Another failure pattern not resolved simply by us- ing a more advanced LLM is Contextual Rela- tionship Blindness. We believe this is partly due to gaps in domain knowledge: during the Section 23 Detail Extraction phase, the model fails to estab- lish the correct relationship between the question and the corresponding sections in a given iteration. As a result, it may overlook crucial information, leading to deficiencies during the information suffi- ciency check stage and causing the system to loop to later sections, ultimately selecting βNone of the Above.β For example, in our Earth Science dataset (specifically, the subfield of urban remote sens- ing), the main study targets are urban landscapes or buildings. Consider question #8 from the batch: Are auxiliary features used beyond raw spectral bands? In remote sensing, it is well understood that the main features are spectral bands and radar bands. Auxiliary features are additional derived features, such as vegetation indices (NDVI, LAI, FAPAR), land surface temperature, albedo, emissivity, topo- graphic information (e.g., DEM), or other external variables. However, when the LLM does not pick up the meaning or understand the term, even if the paragraph includes terms like NDVI, the rela- tionship may still be overlooked. Cases like these highlightingIntraBenchremain challenging. To improve the usability of the dataset for researchers who are not familiar with cross-domain terminol- ogy, we have annotated these terms in the bench- mark; however, these annotations are not provided as inputs to the agent. D Robustness of IntrAgent to Mapping Model and Input Variability Detail D.1 Section Heading Rework In Section 6.4, we investigated howIntrAgentper- forms when provided with subpar or non-standard section headings in scientific literature. Specifi- cally, we examine whether the agent can maintain accuracy even when the section structure is rewrit- ten in unfamiliar or distorted styles. To conduct this study, we extracted all section headings from the paper by Thrift et al. (Thrift and Ragan, 2019), listed in the first column of Ta- ble 20. We then created three alternate versions of these headings: a simplified "Beginner Style", a distorted "Highly Noisy Version", and a stylized "Shakespearean Style", each shown in the subse- quent columns. For each variant, we replaced the original section titles in the document with the newly generated ver- sions.IntrAgentwas then applied to the modified texts using GPT-4o as the backbone LLM and GPT- 4.1 as the mapping model, under default settings. 24 Original TitleBeginner StyleHighly Noisy VersionShakespearean Style A R T I C L E I N F OStuff About This PaperZXCV ARTICLE BLAHPrologue of This Learned Work A B S T R A C TWhat This Thingβs AboutQWERT ABSTRACT THING Abstract of Most Worthy Endeavor 1. Introduction 1. Starting Out1. asdjkh Introduction blah1. An Overture to the Matter at Hand 2. Materials and methods2. Stuff We Used and Did2. jkjhas Materials and doings 2. Of Materials Gathered and Deeds Undertaken 2.1. Bacteria and sample preparation 2.1. Germs and How We Got Them 2.1. ajskdfh Bacteria things and getting stuff 2.1. Of Bacteriumβs Harvest and Preparations Befitting Study 2.2. AFM imaging2.2. Taking Tiny Pics2.2. kjashdf AFM looking-at-things 2.2. Wherein the Art of Atomic Force Is Employed to Gaze Upon the Minuscule 2.3. SERS measurements 2.3. Laser Stuff on the Germs 2.3. asdkljfh SERS zappy laser stuff 2.3. Of Golden Gleams and SERSβThe Spectral Oracle 2.4. Machine learning model development 2.4. Computer Things We Tried 2.4. zmxnv Machine learning magic box 2.4. Wherein the Machine, Apt in Learning, Is Taught to Discern 3. Results3. What Happened3. kjsdhf Results what happened 3. Revelations and Findings Most Curious 3.1. AFM analysis of microbial shape and dimension 3.1. Shapes of the Germs3.1. qwekjh Germ shapes and sizes 3.1. Of Form and MeasureβThe Shape of the Microscopic Realm 3.2. SERS signature of periodontal pathogens 3.2. Shiny Lights from Germs 3.2. kjasdh SERS noise from germs 3.2. Upon the Lightβs Whisper: Signatures of the Mouthβs Hidden Foes 3.3. ML-enabled identification of periodontal pathogens 3.3. Using Computer to Guess Germs 3.3. hgjdksh Computer guessing germ names 3.3. The Thinking Engineβs Triumph in Unmasking the Pathogen 4. Discussion 4. What We Think It Means4. lkjdfh What we think (maybe?) 4. A Discourse Upon the Meaning of These Happenings 5. Conclusion5. Wrapping It Up5. THE END (idk lol)5. A Summation, As the Curtain Draws Near CRediT authorship contribution statement Who Did Whatblorpflorp CRediT who did what maybe Of Quills and Labors: A Testament to Those Who Contributed Declaration of competing interest We Donβt Got Any Fights About This nope-no-fightz Declaration or whatever Of Conflicts, None Declared nor Concealed Data availability Where the Stuff Iscanhasdata Data place??Where Might the Curious Find the Data? AcknowledgementThanks and Shoutoutsthxbye AcknowlegmintsGratitude, in Verse and Spirit Appendix A. Supplementary data Extra Stuff We Didnβt Put Up There Appendix A. More junk we had Appendix the First: Scrolls of Supplement Yet Untold ReferencesBooks and Papers We Looked At Books n Stuff We Readed (aka refs) Tomes and Records ConsultedβA Catalog of the Wise Table 20: Rewriting original section headings from Thrift et al. (Thrift and Ragan, 2019) into three alternative stylesβBeginner Style, Highly Noisy Version, and Shakespearean Styleβto evaluate the robustness ofIntrAgent to variation in document structure. 25 D.2 Section Heading Rework - Highly Noisy Version Reasoning Example This is a detailed reasoning example from the SERS dataset Thrift et al. (Thrift and Ragan, 2019), using the second section title rewrite variation, labeled as the Highly Noisy Version. The rewritten section headings are provided in Table 20. With paper main title provided alongside with the research question for this case is: What are the material and structural properties, or the morphol- ogy, of the SERS substrates used? After the Hierarchy Preservation step, the pro- cessed section headings are as follows: ZZXCV ARTICLE BLAH; QWERT AB- STRACT THING; asdjkh Introduction blah; 2. jkjhas Materials and doings - 2.1. ajskdfh Bacteria things and getting stuff; 2. jkjhas Materials and doings - 2.2. kjashdf AFM looking-at-things; 2. jkjhas Materials and doings - 2.3. asd- kljfh SERS zappy laser stuff; 2. jkjhas Materials and doings - 2.4. zmxnv Ma- chine learning magic box; 3. R kjsdhf Results what happened - 3.1. qwekjh Germ shapes and sizes; 3. R kjsdhf Re- sults what happened - 3.2. kjasdh SERS noise from germs; 3. R kjsdhf Results what happened - 3.3. Mhgjdksh Com- puter guessing germ names; 4. lkjdfh What we think (maybe?); 5. THE END (idk lol); blorpflorp CRediT who did what maybe; nope-no-fightz Declaration or whatever; canhasdata Data place??; thxbye Acknowlegmints; more junk we had; Books n Stuff We Readed (aka refs) In semantic section ranking, we derived the rea- soning step: 1. 2. jkjhas Materials and doings - 2.3. asdkljfh SERS zappy laser stuff β This section specifically mentions "SERS" in its title, indicating it likely contains detailed information about the SERS substrates, including their materials and structure or morphology. 2. 2. jkjhas Materials and doings - 2.1. ajskdfh Bacteria things and getting stuff β As part of the "Materials and doings" main section, this subsection may dis- cuss materials related to the experimen- tal setup, possibly including the SERS substrates. 3. 2. jkjhas Materials and doings - 2.2. kjashdf AFM looking-at-things β This subsection likely involves materials char- acterization using AFM, which may in- clude details about substrate morphology or materials. 4. asdjkh Introduction blah β The in- troduction often contains a summary of the materials and methods used, poten- tially offering an overview of the SERS substrate materials and structure. ... This ultimately leads to correct section alloca- tion, enabling detailed information gathering and passing the information sufficiency check without difficulty. It also results in accurate final mapping to the correct multiple-choice answer. This exam- ple highlights the modelβs ability to extract relevant information from scrambled headings by reasoning through their underlying intent, even in the pres- ence of noise during section ranking. E Prompts Here is the list of prompts we used in our study. E.1 Section Ranking Prompts E.1.1 Hierarchy Preservation Prompt Prompt: Hierarchy Preservation hierarchy_preservation_prompt = """ You are an expert in scientific research and academic writing, proficient in analyzing the structural organization of research papers. Your task is to infer the hierarchical structure of a research paper based on a list of section titles provided in sequential order. ,β ,β ,β ,β ,β ,β ,β --- ### **Task** You are given a list of section titles extracted from a research paper, shown in the order they appear in the document: ,β ,β ,β headings These titles are not annotated to indicate their level (e.g., section, subsection, sub-subsection). Your job is to infer the **tree structure** by determining which titles are: ,β ,β ,β ,β - **Main sections** 26 - **Subsections** under a main section - **Sub-subsections** under a subsection --- ### **Rules** - Maintain the **original order** of the titles.,β - Do **not** introduce any new titles or remove existing ones.,β - Use the **first title** as the likely **main title** of the paper, unless it's clearly a preamble (e.g., "Abstract", "Introduction"). ,β ,β ,β - Infer nesting levels based on typical academic paper structure and semantic cues in the titles. ,β ,β - Structure the output using this format: - `"Section Title"` (main section) - `"Section Title - Subsection Title"` (subsection),β - `"Section Title - Subsection Title - Sub-subsection Title"` (sub-subsection) ,β ,β --- ### **Example** Given: ["Experiment setup", "sample preparation", "device setup", "SERS measurements", "Result and Discussion"] ,β ,β Expected output: ["Experiment setup", "Experiment setup - sample preparation", "Experiment setup - device setup", "Experiment setup - SERS measurements", "Result and Discussion"] --- ### **Expected Output Format** - Provide the inferred tree structure in a Python list format.,β - Follow the form: `["Section", "Section - Subsection", "Section - Subsection - Sub-subsection", ...]` ,β ,β - Output **only the list**. Do not include explanations, reasoning steps, or extra comments. ,β ,β --- Let's think step by step. """ E.1.2 Reasoning-Based Ranking Prompt Prompt: Reasoning-Based Ranking ranking_prompt = """ You are an AI research assistant with expertise in analyzing academic papers. Your task is to determine which sections of a paper are most likely to contain the answer to a given research question. The section titles are provided in a **tree structure**, where main sections and their subsections are denoted using a hyphen ("-"). Your goal is to rank these sections and subsections from **most relevant to least relevant** in relation to the research question. ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β --- **Paper Title**: main_title_heading **Research Question**: question **Section Titles**: formatted_headings --- ### **Instructions** - Analyze the research question and the structure of the section titles.,β - Rank the section titles in order of likelihood to contain the answer β from most to least relevant. ,β ,β - Subsections may be more specific than main sections; use their nesting to guide your ranking. ,β ,β - For each section in the ranking, provide a **one-sentence explanation** for its placement. ,β ,β - Maintain the original order of section titles when relevance is tied.,β - Do **not** skip or omit any section titles from the list.,β - Format the final ranking using only integer indices, enclosed in <<<#>>> brackets, based on the position of each title in the provided list (starting from 1). ,β ,β ,β ,β - Do not repeat section titles or include any extra commentary after the final ranking. ,β ,β --- ### **Response Template** Reasoning steps: 1. [Title 1] β [1-sentence justification] 2. [Title 2] β [1-sentence justification] ... **Final ranking:** <<<2>>>, <<<1>>>, <<<3>>>, ... <<<n>>>,β Let's think step by step. """ 27 E.2 Iterative Reading Prompts E.2.1 Action Loop Prompt Prompt: Action Loop main_prompt = """ You are an assistant designed to select the next Action based on the current observation. ,β ,β Each observation contains: - The current chunk index, indicating which chunk of the document you are working on. ,β ,β - The total number of available chunks. - A record of past Actions taken in previous iterations.,β - If the most recent Action was EVALUATE, the observation will also include the evaluation result. ,β ,β --- To ensure accuracy, you must follow these instructions:,β main_prompt_instructions --- Descriptions of all allowed actions: - GET_CHUNK: Retrieve the next Knowledge Chunk. Use this when no chunk is currently loaded or when advancing to the next chunk. ,β ,β ,β - GET_DETAIL: Extract relevant details with respect to the Research Question from the currently loaded Knowledge Chunk. ,β ,β ,β - EVALUATE: Determine whether the extracted details from the current Knowledge Chunk are sufficient to answer the Research Question. ,β ,β ,β - TERMINATE: End the process once you have gathered enough relevant information to confidently answer the Research Question. (TERMINATE is only valid if the most recent evaluation yielded a βYESβ result.) ,β ,β ,β ,β ,β --- Here are some examples: - Past Action Taken: GET_CHUNK. I have already obtained the chunk, so I will GET_DETAIL. ,β ,β - Past Action Taken: GET_CHUNK, GET_DETAIL. then I will now EVALUATE the summary.,β - Past Action Taken: GET_CHUNK, GET_DETAIL, EVALUATE. The result does not seem to be going well, so I will GET_CHUNK again. ,β ,β ,β - Past Action Taken: GET_CHUNK, GET_DETAIL, EVALUATE. The result seems to be going well, so I will TERMINATE the process. ,β ,β - Past Action Taken: GET_CHUNK, GET_DETAIL, EVALUATE, ... ,GET_CHUNK, GET_DETAIL, EVALUATE. This is chunk 10 out of 10. I was supposed to get more chunks, but I have already retrieved the last one. No more knowledge chunks are available, so I will TERMINATE the process. ,β ,β ,β ,β ,β ,β --- Observation: You are currently at chunk current_chunk_index out of total_chunks_len chunks. ,β ,β Past actions taken: past_action evaluation_result --- Response Format: Your final answer must follow the exact format below. Do not include any text outside this format. ,β ,β Format: Reasoning Steps: [In one sentence, explain why this Action is the best next step.],β Action: [Choose one of the following: GET_CHUNK, GET_DETAIL, EVALUATE, TERMINATE] ,β ,β """ E.2.2 Action Loop - Default Balanced Instruction For the instruction given in the loop above, depend on confidence level, different prompt was provided. This given Instruction subject to confidence level 2. This instruction is being used for default setting. Prompt: Action Loop main_prompt_instructions_confidence_level_2 = """,β 1. You will be provided with a list of Knowledge Chunks, ordered by relevance to the research question. Earlier Knowledge Chunks are more likely to contain useful information. ,β ,β ,β ,β 2. Your task is to iteratively retrieve a Knowledge Chunk, extract relevant details, and evaluate whether you can answer the research question. Repeat this process until you can confidently terminate. ,β ,β ,β ,β ,β 3. You will be given a list of predefined actions to select from.,β 4. In each iteration, you must select exactly one action based on the current observation. ,β ,β 28 5. The observation includes: the current Knowledge Chunk index, the total number of available Knowledge Chunks, and the list of past actions taken. ,β ,β ,β 6. The available actions are: GET_CHUNK, GET_DETAIL, EVALUATE, and TERMINATE, as defined in the action list. ,β ,β 7. If the most recent action in the past actions taken was GET_CHUNK, your next action **must** be GET_DETAIL to extract information from the current Knowledge Chunk β do not perform this extraction yourself. ,β ,β ,β ,β ,β 8. If the most recent action was GET_DETAIL, your next action **must** be EVALUATE to assess whether the gathered details are sufficient to answer the research question β do not perform the evaluation yourself. ,β ,β ,β ,β ,β 9. You must select only one action at a time β never choose multiple actions in a single step. ,β ,β 10. Since Knowledge Chunks are ordered by relevance to the research question, earlier Knowledge Chunks are more likely to contain useful information. Use this ordering to guide your reading sequence. ,β ,β ,β ,β ,β 11. You must TERMINATE the process once you have gathered enough information to confidently answer the research question. ,β ,β ,β 12. Pay attention to the total number of Knowledge Chunks. If your most recent EVALUATE action was performed on the **last available Knowledge Chunk** and the result was insufficient, you must TERMINATE the process due to lack of additional information. ,β ,β ,β ,β ,β ,β 13. The expected workflow is: GET_CHUNK β GET_DETAIL β EVALUATE β TERMINATE (if evaluation result is sufficient). If not, continue to the next Knowledge Chunk. If you reach the final Knowledge Chunk and the evaluation is still insufficient, you must TERMINATE. ,β ,β ,β ,β ,β ,β ,β 14. You may TERMINATE early if an EVALUATE action confirms that the gathered information is sufficient to answer the research question with high confidence. ,β ,β ,β ,β 15. After your reasoning, output your response in the specified format.,β """ E.2.3 Section Detail Extraction Prompt Prompt: Action: Section Detail Extraction get_detail_prompt = """ You are a research assistant helping extract detailed information relevant to the given Research Question based on a Knowledge Chunk given. ,β ,β ,β --- Research Question: question Knowledge Chunk: chunk --- Task: - Extract all key points from the current Knowledge Chunk **only as they relate to the Research Question**. ,β ,β - Include all relevant information such as scientific terms, numerical data, experimental results, measurements, statistical indicators, conclusions, and any comparative or causal statements that explicitly support or address the Research Question. ,β ,β ,β ,β ,β ,β - Do not infer or assume missing content. If essential information is absent, clearly state what is missingβavoid speculation or completion. ,β ,β ,β - When a sentence directly answers or supports the Research Question, quote it **verbatim** from the **original chunk**. ,β ,β ,β - If there are multiple key points, present them as a structured list in bullet point format. ,β ,β - If there are no key points relevant to the Research Question, clearly state: *"This chunk does not provide relevant information."* ,β ,β ,β - Perform this task step by step, ensuring that each extracted detail is justified and directly traceable to the original content. ,β ,β ,β --- Response Format: Your final answer must follow the exact format below. Ensure strict adherence to this structure. Do not include additional explanations outside of this format. ,β ,β ,β ,β Format: Reasoning Steps: [reasoning step by step goes here],β Details: The Section_title is <Section_title>.[Details based on the Research Question and chunk go here. Include all relevant findings and quote supporting sentences, present them as a structured list.] ,β ,β ,β ,β ,β """ 29 E.2.4 Default Balanced Information Sufficiency Check Prompt The Information Sufficiency Check Prompt, con- fidence level 2. This prompt is being used for the default setting. Prompt: Sufficiency Check Confidence Level 2 evaluation_prompt_confidence_level_2 = """ You are a research assistant tasked with evaluating whether the provided details are both sufficient and accurate to answer a given research question. Based solely on the current details, determine whether they contain all the necessary and correct information required to answer the research question. Your evaluation must include direct references to original sentences from the provided details to support your reasoning. ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β --- Research Question: question Current Details: observation_stage --- Task: - Assess whether the current details contain all essential and accurate information needed to answer the Research Question. ,β ,β ,β - Respond with "YES" **only if** the provided details fully and correctly address the research question with no missing elements or uncertainties. ,β ,β ,β - In this case, provide a complete answer supported by quoted or clearly referenced content from the current Knowledge Chunk. ,β ,β ,β - If the information is incomplete, partially correct, or uncertain in any way, respond with "NO." ,β ,β - Clearly identify what specific information is missing or ambiguous.,β - Then provide the closest possible answer using only the available details. ,β ,β - If no relevant information is present at all, write: *"No information in given details."* ,β ,β - After reasoning step by step, output both `Sufficiency` and `Detail_Answer`. Your `Detail_Answer` must include direct evidence from the knowledge chunk β avoid general summaries. ,β ,β ,β ,β ,β --- Response Format: Your response must strictly follow the format below:,β Sufficiency: [YES or NO] Detail_Answer: The Section_title is <Section_title>. [Provide your best possible answer to the research question, using direct references from the details. Explain your reasoning step by step, ensuring each claim is supported by the provided content.] ,β ,β ,β ,β ,β ,β """ E.2.5 Detail Aggregation Final Answer Prompt Prompt: Summary for Final Answer full_set_answer_prompt = """ You are a research assistant tasked with synthesizing a final answer to the research question based on the evaluations of multiple knowledge chunks provided below. Each evaluation entry includes: ,β ,β ,β ,β ,β - A sufficiency judgment (YES or NO), - A detailed answer (Detail_Answer), - Referenced original sentences, each tagged with its Chunk number and Section title. ,β ,β Use the complete set of evaluation entries to construct a coherent, well-supported, and evidence-based final answer to the research question. You must include direct references to the **original sentences**, along with their corresponding **Chunk number** and **Section title**, to justify each claim you make. ,β ,β ,β ,β ,β ,β ,β ,β --- Research Question: question Evaluation Entries: observation_stage --- Task: - Synthesize the provided evaluation entries to generate a comprehensive and conclusive answer to the Research Question. ,β ,β ,β - Your answer must be fully supported by specific content from the evaluation entries. ,β ,β - When making a claim, **explicitly cite** the source using the format: `"Quoted sentence from the original text" (Chunk #, Section_title: <Section_title>)`. ,β ,β ,β ,β - If multiple entries support the same point, cite each one.,β 30 - Do **not** introduce new information or inferred content that is not present in the evaluation entries. ,β ,β - Avoid vague generalizations β every component of the answer must be evidence-backed. ,β ,β - After reasoning step by step, output `Final_Answer` using the exact format below. ,β ,β --- Response Format: Final_Answer: [Provide your final answer with supporting evidence from the evaluation entries. For every claim, cite the original sentences along with the Chunk number and Section title they came from. Explain your reasoning step by step, covering all required aspects of the Research Question.] ,β ,β ,β ,β ,β ,β ,β """ E.3 Evaluation Protocol Prompts E.3.1 Evaluation Protocol - Baseline Final Answer Prompt Prompt: Final Answer - Choices Mapping full_set_answer_prompt = """ You are a research assistant tasked with synthesizing a final answer to the research question based on the evaluations of multiple knowledge chunks provided below. Each evaluation entry includes: ,β ,β ,β ,β ,β - A sufficiency judgment (YES or NO), - A detailed answer (Detail_Answer), - Referenced original sentences, each tagged with its Chunk number and Section title. ,β ,β Use the complete set of evaluation entries to construct a coherent, well-supported, and evidence-based final answer to the research question. You must include direct references to the **original sentences**, along with their corresponding **Chunk number** and **Section title**, to justify each claim you make. ,β ,β ,β ,β ,β ,β ,β ,β --- Research Question: question Evaluation Entries: observation_stage --- Task: - Synthesize the provided evaluation entries to generate a comprehensive and conclusive answer to the Research Question. ,β ,β ,β - Your answer must be fully supported by specific content from the evaluation entries. ,β ,β - When making a claim, **explicitly cite** the source using the format: `"Quoted sentence from the original text" (Chunk #, Section_title: <Section_title>)`. ,β ,β ,β ,β - If multiple entries support the same point, cite each one.,β - Do **not** introduce new information or inferred content that is not present in the evaluation entries. ,β ,β - Avoid vague generalizations β every component of the answer must be evidence-backed. ,β ,β - After reasoning step by step, output `Final_Answer` using the exact format below. ,β ,β --- Response Format: Final_Answer: [Provide your final answer with supporting evidence from the evaluation entries. For every claim, cite the original sentences along with the Chunk number and Section title they came from. Explain your reasoning step by step, covering all required aspects of the Research Question.] ,β ,β ,β ,β ,β ,β ,β """ E.3.2 Evaluation Protocol - Final Answer to MCQ Answer Mapping Prompt: Final Answer - Final Answer to MCQ Answer Mapping prompt = f""" You are a research assistant. Your task is to map a short answer to a research question, derived from a scientific research paper, to the most appropriate option among the provided answer choices in a multiple-choice question (MCQ) format. ,β ,β ,β ,β ,β ,β We are given the following research question:,β question To support your understanding of the research question, here is some relevant contextual information: ,β ,β additional_note Here are the extracted answers to the research question, taken from the research paper: ,β ,β relevant_chunks Here are the answer choices: paper_choice **Instructions:** 31 1. You must select **only one answer choice** from the list that is either correct or most relevant to the research question. ,β ,β ,β 2. Provide your answer as a single letter (e.g., A, B, C, D, E, F) enclosed in the format <<<ANS>>>. Do not include any other text in your final output. ,β ,β ,β 3. You may include step-by-step reasoning to justify your decision, but your final output must consist of only one answer choice in the required format. ,β ,β ,β 4. This is a science and technology-related research question. For cases where the short answer involves numerical values: ,β ,β ,β - You must map the short answer to the most accurate and relevant answer choice. ,β ,β - If the short answer and an answer choice represent numerical values, apply the following decimal precision alignment rules: ,β ,β ,β - If the short answer and an answer choice have the **same number of decimal places**, compare them directly without rounding. ,β ,β ,β ,β - If they have **different numbers of decimal places**, round the value with **more digits** to match the **shorter decimal precision**, then compare. ,β ,β ,β ,β - After alignment, select the answer choice that **exactly matches** the short answer. ,β ,β - If no answer choice matches after applying these rules, select **βF. None of the above.β** ,β ,β - **Examples**: - Short answer = 1.1111, answer choices include 1.1 β round short answer to 1.1 β match β select 1.1 ,β ,β ,β - Short answer = 2.01, answer choices include 1.98 β both have two decimal places β no rounding β no match β select βF. None of the aboveβ ,β ,β ,β ,β 5. Never provide multiple answers. **Example of output format:** *reasoning steps* --- <<<C>>> Let's think step by step. """ E.4Confidence Level Ablation Study Prompts E.4.1 Confidence Level 1 Ablation Study - Conservative Information Sufficiency Check Prompt Prompt: Sufficiency Check Confidence Level 1 evaluation_prompt_confidence_level_1 = """ You are a research assistant tasked with evaluating whether the provided details are both sufficient and accurate to answer a given research question. Based solely on the current details, determine whether they contain all the necessary and correct information required to answer the research question. You must act conservatively β your evaluation should follow the strictest standard: any ambiguity, missing element, or lack of clarity must result in a "NO". ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β --- Research Question: question Current Details: observation_stage --- Task: - Assess whether the extracted details contain all **necessary elements** to fully and unambiguously answer the Research Question. ,β ,β ,β - Only respond with "YES" if **all aspects** of the research question are addressed with precise and complete evidence from the provided knowledge chunk. ,β ,β ,β ,β - No assumptions, inferred logic, or external knowledge are allowed.,β - If **any part** of the required information is missing, vague, incomplete, or unclear, respond with "NO". ,β ,β ,β - Clearly identify what is missing or uncertain.,β - Then provide the closest possible answer that can be supported solely by the available content from the Knowledge Chunk. ,β ,β ,β - If no relevant information is present at all, state: *"No information in given details."* ,β ,β - After step-by-step reasoning, output both `Sufficiency` and `Detail_Answer`. Your `Detail_Answer` must include direct evidence from the knowledge chunk β avoid general summaries. ,β ,β ,β ,β ,β --- Response Format: 32 Your response must strictly follow the format below:,β Sufficiency: [YES or NO] Detail_Answer: The Section_title is <Section_title>. [Provide your best possible answer to the research question, using direct references from the details. Explain your reasoning step by step, addressing all required components.] ,β ,β ,β ,β ,β ,β """ E.4.2 Confidence Level 3 Ablation Study - Aggressive Information Sufficiency Check Prompt Prompt: Sufficiency Check Confidence Level 3 evaluation_prompt_confidence_level_3 = """ You are a research assistant tasked with evaluating whether the provided details are both sufficient and accurate to answer a given research question. You may act more aggressively and confidently. Based solely on the current details, determine whether they contain **enough relevant content to reasonably answer** the research question, even if some minor points are not fully explicit. ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β ,β --- Research Question: question Current Details: observation_stage --- Task: - Assess whether the current details provide **most or all of the key information** needed to reasonably answer the Research Question. ,β ,β ,β - Respond with "YES" if the answer can be supported using the provided information, even if a few supporting details are missing, implicit, or only partially clear. ,β ,β ,β ,β - If confident, provide a full answer backed by the best available references from the Knowledge Chunk. ,β ,β - Only respond with "NO" if **critical** information is missing or the answer would be too speculative. ,β ,β - Clearly explain what is missing or unclear.,β - Then provide the closest possible answer based only on the available content. ,β ,β - If no relevant information is present at all, state: *"No information in given details."* ,β ,β - After step-by-step reasoning, output both `Sufficiency` and `Detail_Answer`. Your `Detail_Answer` must include direct evidence from the knowledge chunk β avoid general summaries. ,β ,β ,β ,β ,β --- Response Format: Your response must strictly follow the format below:,β Sufficiency: [YES or NO] Detail_Answer: The Section_title is <Section_title>. [Provide your best possible answer to the research question, using direct references from the details. Explain your reasoning step by step, addressing all required components.] ,β ,β ,β ,β ,β ,β """ E.4.3 Confidence Level 1 Ablation Study - Action Loop Conservative Instruction Prompt: Confidence Level Case Study Prompts 1 main_prompt_instructions_confidence_level_1 = """,β 1. You will be provided with a list of Knowledge Chunks, ordered by relevance to the research question. Earlier Knowledge Chunks are more likely to contain useful information. ,β ,β ,β ,β 2. Your task is to iteratively retrieve a Knowledge Chunk, extract relevant details, and evaluate whether you can answer the research question. Repeat this process until you have evaluated all available Knowledge Chunks or gathered enough information to confidently terminate. ,β ,β ,β ,β ,β ,β ,β 3. You will be given a list of predefined actions to select from.,β 4. In each iteration, you must select exactly one action based on the current observation. ,β ,β 5. The observation includes: the current Knowledge Chunk index, the total number of available Knowledge Chunks, and the list of past actions taken. ,β ,β ,β 6. The available actions are: GET_CHUNK, GET_DETAIL, EVALUATE, and TERMINATE, as defined in the action list. ,β ,β 7. If the most recent action in the past actions taken was GET_CHUNK, your next action **must** be GET_DETAIL to extract information from the current Knowledge Chunk β do not perform this extraction yourself. ,β ,β ,β ,β ,β 33 8. If the most recent action was GET_DETAIL, your next action **must** be EVALUATE to assess whether the gathered details are sufficient to answer the research question β do not perform the evaluation yourself. ,β ,β ,β ,β ,β 9. You must select only one action at a time β never choose multiple actions in a single step. ,β ,β 10. Since Knowledge Chunks are ordered by relevance to the research question, earlier Knowledge Chunks are more likely to contain useful information. Use this ordering to guide your reading sequence. ,β ,β ,β ,β ,β 11. You must TERMINATE the process only after either (a) all available Knowledge Chunks have been evaluated, or (b) you are confident that the research question can be answered based on the most recent evaluation. ,β ,β ,β ,β ,β 12. Pay attention to the total number of Knowledge Chunks. If your most recent EVALUATE action was performed on the **last available Knowledge Chunk**, you must TERMINATE the process regardless of the outcome. ,β ,β ,β ,β ,β 13. The expected workflow is: GET_CHUNK β GET_DETAIL β EVALUATE β TERMINATE (if evaluation result is sufficient). If not, continue to the next chunk. If you reach the final chunk and the evaluation is still insufficient, you must TERMINATE. ,β ,β ,β ,β ,β ,β 14. Do not TERMINATE early based solely on assumptions about relevance. Always continue until a conclusive evaluation result is available or all chunks are exhausted. ,β ,β ,β ,β 15. After your reasoning, output your response in the specified format.,β """ E.4.4 Confidence Level 3 Ablation Study - Action Loop Aggressive Instruction Prompt: Confidence Level Case Study Prompts 3 main_prompt_instructions_confidence_level_3 = """,β 1. You will be provided with a list of knowledge chunks, ordered by relevance to the research question. Earlier chunks are more likely to contain useful information. ,β ,β ,β ,β 2. Your task is to iteratively retrieve a knowledge chunk, extract relevant details, and evaluate whether you can answer the research question. Repeat this process until you can confidently terminate. ,β ,β ,β ,β ,β 3. You will be given a list of predefined actions to select from.,β 4. In each iteration, you must select exactly one action based on the current observation. ,β ,β 5. The observation includes: the current chunk index, the total number of available chunks, and the list of past actions taken. ,β ,β ,β 6. The available actions are: GET_CHUNK, GET_DETAIL, EVALUATE, and TERMINATE, as defined in the action list. ,β ,β 7. If the most recent action in the past actions taken was GET_CHUNK, your next action **must** be GET_DETAIL to extract information from the current knowledge chunk β do not perform this extraction yourself. ,β ,β ,β ,β ,β 8. If the most recent action was GET_DETAIL, your next action **must** be EVALUATE to assess whether the gathered details are sufficient to answer the research question β do not perform the evaluation yourself. ,β ,β ,β ,β ,β 9. You must select only one action at a time β never choose multiple actions in a single step. ,β ,β 10. Since knowledge chunks are ordered by relevance to the research question, earlier chunks are more likely to contain useful information. Use this ordering to guide your reading sequence. ,β ,β ,β ,β ,β 11. You may TERMINATE the process as soon as you believe that additional knowledge chunks are unlikely to provide significantly better information, even if you are not fully confident. ,β ,β ,β ,β ,β 12. If your most recent EVALUATE action was performed on the **last available chunk**, you must TERMINATE the process regardless of the outcome. ,β ,β ,β 13. The expected workflow is: GET_CHUNK β GET_DETAIL β EVALUATE β TERMINATE. You may also TERMINATE early based on the assumption that remaining chunks have diminishing relevance. ,β ,β ,β ,β 14. Terminate aggressively if your evaluation suggests NO from continuing, even if high confidence has not yet been achieved. ,β ,β ,β 15. After your reasoning, output your response in the specified format.,β """ E.5 Contextual RAG Prompt Prompt: Contextual RAG Prompt prompt = """ You are an expert technical writer specialised in contextual retrieval.,β <document> whole_document </document> Here is the chunk we want to situate within the whole document,β <chunk> chunk </chunk> 34 Please give a short succinct context to situate this chunk within,β the overall document for the purposes of improving search retrieval,β of the chunk. Answer only with the succinct context and nothing else.,β """ F Baseline Evaluation Details F.1 Vanilla RAG Vanilla RAG refers to the original implementation proposed by Lewis et al. (Lewis et al., 2020). In our setup, we first convert the PDF versions of pa- pers into plain text. Subsequently, we tokenize the text into fixed-length chunks of 500 tokens with an overlap of 50 tokens between consecutive chunks. We select 500 tokens as it provides a balance be- tween semantic coherence and memory efficiency, allowing enough context to be preserved within each chunk. The 50-token overlap ensures that important information occurring at chunk bound- aries is not lost, improving retrieval continuity and robustness. For embedding generation and similarity compu- tation, we use theall-MiniLM-L6-v2304model to encode each chunk. During retrieval, we adopt a top-kstrategy, defaulting to the top 3 chunks most similar to the given research question, based on co- sine similarity scores. This top-kselection aligns with our evaluation protocol, where these top 3 chunks are passed into the generation module to produce short-form answers. For final answer generation, we evaluate selected backbone LLMs under identical input conditions to ensure a fair comparison. Each model generates an answer conditioned on the retrieved top-3 chunks and the research question. F.2 Contexual RAG To enhance retrieval precision beyond conventional dense embedding approaches, we implemented a Contextual Retrieval-Augmented Generation (Con- textual RAG) framework (Anthropic, 2024). Un- like vanilla RAG, which embeds document chunks in isolation, Contextual RAG explicitly augments each chunk with a succinct, model-generated con- text situating it within the broader document struc- ture. This implementation effectively bridges se- mantic and structural understanding, allowing the retrieval component to capture both what is being discussed and where it fits contextually within the paper. This contextual enrichment helps disam- biguate semantically similar fragments, improv- ing the embedding modelβs ability to capture fine- grained relationships between queries and relevant content. Our pipeline first extracts raw text given litera- ture, followed by cleaning and token-level chunk- ing with a 500-token window and 50-token over- lap to preserve local continuity. For each chunk, a lightweight contextualization module built on GPT-4o generates concise summaries that describe how the chunk fits within the overall document; prompt used see E.5. These augmented chunks are then encoded using modelsβeither E5-Mistral-7B- Instruct or GritLM-7Bβ for efficient batch infer- ence. Query and chunk embeddings are computed with cosine similarity to identify the top-k most relevant text segments per question, enabling high- precision retrieval aligned with the documentβs nar- rative structure. F.3 DRAGIN We adopt DRAGIN (Dynamic Retrieval Aug- mented Generation based on the real-time Informa- tion Needs of LLMs)(Su et al., 2024b) as the base- line framework for dynamic retrieval-augmented generation. DRAGIN explicitly determines both when to retrieve and what to retrieve based on the modelβs internal uncertainty and token-level salience signals, enabling fine-grained control over evidence acquisition during generation. To adapt DRAGIN to our customized scenario involving a small-scale paragraph-level corpus and chain-of-thought-style few-shot prompting, we in- troduce the following modifications: β’ Paragraph-Level Corpus Construction. We replace the default Wikipedia corpus (psgs_w100.tsv) used in the official DRA- GIN release with a custom paragraph-level corpus (DRAGIN_paragraphs.tsv) tailored to our domain. Each entry consists of a unique paragraph ID, a document title, and the para- graph text. This design enables the retrieval module to operate at a finer granularity, which is essential for scenarios involving a small collection of five domain-specific documents. β’Few-Shot Prompting with Summary- Driven CoT Chains.We generate ten chain-of-thought (CoT) reasoning exemplars from the five domain-specific documents. For each test instance,IntrAgentsummarizes 35 the top five retrieved paragraphs to construct a condensed context representation. These summaries serve as inputs to guide CoT generation and are incorporated into thedemo field during theinference()stage.For example: βThe spatial extent of the study area is approximately 966 kmΒ², de- rived from the Residential cate- goryβs area (398 kmΒ²) and its pro- portion (41.20%) of the total imper- vious surface...β This procedure ensures that the few-shot prompt incorporates domain-relevant knowl- edge in a structured format, strengthening the alignment between the modelβs reasoning path and the underlying evidence. β’Dynamic Retrieval at Inference Time. Dur- ing inference, DRAGIN actively monitors for hallucination-prone segments and triggers retrieval of relevant evidence in real time. Retrieved paragraphs are incrementally inte- grated into the generation context, enabling iterative refinement of the modelβs output. Un- like short-form QA systems, DRAGIN pro- duces complete chain-of-thought reasoning sequences, along with auxiliary metadata such as retrieval count and token usage for diagnos- tic analysis. Following CoT generation, we initiate a final synthesis stage. The full reasoning trace, rather than the retrieved paragraphs themselves, is incor- porated into a standardized prompt alongside the original question. This prompt is then passed to an external language model (e.g., GPT-4-turbo, GPT- 4o, or DeepSeek-Chat), which is instructed to sum- marize the reasoning and produce a concise final answer. This decoupled architecture ensures inter- pretability during intermediate steps while main- taining fluency and coherence in the final output. F.4 R 2 AG We adopt R 2 AG (Reranking-augmented Retrieval- Augmented Generation) as one of our baseline(Ye et al., 2024). R 2 AG enhances generation quality by incorporating a reranking module that explicitly guides evidence selection, thus mitigating retrieval noise and improving relevance alignment in the final output. To implement the R 2 AG pipeline in our setting, we follow the procedures below: Retriever Fine-Tuning and Dense Indexing. We fine-tune a Sentence-BERT retriever using posi- tive and negative question-paragraph pairs. Specifi- cally, we leverage the DuReader benchmark, which provides human-annotated relevance labels for question-passage pairs, ensuring high-quality su- pervision during training. Positive samples are pas- sages annotated as relevant to the given question, while negatives are randomly selected or chosen as hard negatives from the same context. After train- ing, the retriever encodes all corpus paragraphs into dense embeddings, which are then indexed using FAISS (IndexFlatIP) for efficient top-ksimilarity search during retrieval. Construction ofR 2 AG Training Samples. Each training instance comprises a question, reference answer,kretrieved paragraph embeddings, and a binary relevance label for each paragraph. Labels are heuristically assigned based on similarity with the gold answer using token overlap or ROUGE. These samples are used to jointly supervise both generation and reranking. Relevance-Aware Generation via RFormer. During training, paragraph embeddings are pro- cessed by R-Former, a lightweight transformer en- coder that outputs a relevance-guided vector. This vector is projected to match the hidden dimension of a frozen LLM (Here we used LLaMA) and in- jected into the prompt via a placeholder token. The model jointly optimizes generation loss (over the answer) and binary classification loss (over para- graph relevance). Inference At inference time, R 2 AG retrieves top- kpassages, processes them via RFormer, and in- jects the reranking-aware signal into the prompt before decoding with the LLM. F.5 LongRAG We adopt LongRAG (Jiang et al., 2024) as one of our baselines. LongRAG enhances retrieval- augmented reasoning through a dual-perspective design consisting of a Hybrid Retriever for coarse- to-fine semantic recall, an LLM-augmented Infor- mation Extractor for global context recovery, and a CoT-guided Filter for relevance refinement. To adapt LongRAG to our scientific literature corpus and domain-specific QA task, we apply several modifications. First, each paper is segmented into seman- tically coherent chunks of approximately 200 36 words, using sentence boundaries as the min- imal division unit.Each chunk is encoded with theintfloat/multilingual-e5-large dual encoder and stored in a FAISS in- dex for dense retrieval.In the fine- grained re-ranking stage, the cross-encoder nreimers/mmarco-mMiniLMv2-L12-H384-v1is used with default settings(chunk_size = 200, top-k = 7). Second, the retrieved chunks are mapped back to their original paragraphs to restore paragraph- level semantic continuity. When multiple chunks correspond to the same paragraph, only the most relevant one is retained. The resulting paragraphs are passed to an LLM-based extractor, which gen- erates a global contextual representation capturing structural and semantic relationships across the pa- per. Third, we retain LongRAGβs two-stage CoT- guided filtering mechanism to eliminate redundant or irrelevant content. The model produces interme- diate reasoning traces to identify evidence-bearing segments and synthesizes an intermediate answer by combining globally summarized and filtered in- formation. In our adaptation, GPT-3.5-Turbo-16k serves as the backbone model for extraction, filter- ing, and generation. Finally, a more capable external LLMβsuch as GPT-4o, GPT-4.1, DeepSeek-R1, Gemini 2.5 Pro, o3, o4-mini, or Llama-3.1-70B-Instruct-Turboβis employed to refine intermediate outputs and pro- duce the final synthesized answers. Accuracy is reported as the primary evaluation metric, consis- tent with our multiple-choice benchmark protocol. The models used in each stage are summarized in Table 21. F.6 Agent LUMOS This section details our customized version of the LUMOS iterative framework(Yin et al., 2024), adapted specifically for the multiple-choice ques- tion (MCQ) task. Unlike the original LUMOS framework designed for open-domain QA, our ver- sion retains its iterative structure with separate Planning, Grounding, and Execution modules. At each step, the system generates subgoals, for- mulates paragraph retrieval actions, retrieves rel- evant evidence, and accumulates results. Once enough evidence is collected, we use an external LLM (GPT-4o, GPT-4-turbo, or DeepSeek-Chat) to synthesize the final answer based on all retrieved content. Our method retains LUMOSβs modular design, consisting of a planning module, grounding mod- ule, and execution module. The models used in each stage are summarized in Table 22: The baseline system adopts an iterative multi- stage reasoning framework composed of four core modules: planning, grounding, execution, and final synthesis. The process begins by initializing an interme- diate result dictionary with the original question and an initial context. In each iteration, the plan- ning module generates a reasoning subgoal based on the question, accumulated results, and previous actions. This is achieved using a pretrained model, lumos_complex_qa_plan_iterative, which is built on LLaMA 2β7B. The iteration process ter- minates either when the planning module outputs an explicit stop signal (e.g., βYes, I will stop plan- ning.β) or when a fixed maximum of 5 steps is reached. Once a subgoal is generated, the system proceeds to the grounding stage, where the subgoal is trans- lated into a structured paragraph retrieval action conditioned on the current reasoning context. The execution module is responsible for retriev- ing the most relevant paragraph. It segments the context into overlapping chunks and selects the top match based on semantic similarity using a custom retriever. To evaluate the effect of evidence gran- ularity, we test retrieval based on the top-5, top-7, and top-9 most relevant paragraphs. After sufficient evidence has been collected through iterations, the system enters the final syn- thesis stage. All retrieved content is compiled into a standardized prompt, including the original ques- tion and numbered evidence chunks. This prompt is then passed to an external large language modelβ such as GPT-4-turbo, GPT-4o, or DeepSeek-Chatβ to generate the final answer. The model is in- structed to synthesize an answer strictly based on the retrieved content, ensuring transparency and faithfulness. This modular design supports step-wise and in- terpretable reasoning, separates control logic from retrieval, and uses the LLM exclusively in the final answer synthesis stage. F.7 PaperQA2 We reproduce and adapt the PaperQA2 frame- work for our controlled setting involving a small fixed corpus of input PDFs(Skarlinski et al., 2024), rather than a large open-domain scientific literature 37 StageModel / Tool Dual-Encoder Retrieval intfloat/multilingual-e5-large Cross-Encoder Re-ranking nreimers/mmarco-mMiniLMv2-L12-H384-v1 Information Extraction (Global) GPT-3.5-Turbo-16k CoT-guided Filtering (Local) GPT-3.5-Turbo-16k Answer Generation (Intermediate) GPT-3.5-Turbo-16k Final Answer External LLMs:GPT-4o,GPT-4.1,DeepSeek-R1,Gemini 2.5 Pro,o3, o4-mini, Llama-3.1-70B-Instruct-Turbo Table 21: Models used in each stage of the LongRAG baseline. ModuleRoleModel Used PlanningGenerate reasoning subgoals LLaMA 2β7B GroundingConvert subgoals into retrieval actions / ExecutionExecute retrieval over provided context dpr-reader-multiset-base Final AnswerSynthesize final answer using retrieved content External LLM (GPT-4-Turbo, GPT-4o, or DeepSeek-Chat) Table 22: Models used in each stage of the LUMOS baseline. GPT-4oGPT-4.1DeepSeek-R1 PhysPHESAvg. PhysPHESAvg. PhysPHESAvg. LUMOS (3 chunk)43.1 36.4 41.7 40.4 41.5 41.8 40.0 41.1 38.5 40.0 38.3 38.9 LUMOS (5 chunk β )52.3 43.6 46.7 47.5 50.8 47.3 51.7 49.9 50.8 49.1 46.7 48.9 LUMOS (7 chunk)58.5 49.1 51.7 53.1 56.9 52.7 55.0 54.9 55.4 56.4 51.7 54.5 LUMOS (9 chunk)61.5 52.7 53.3 55.8 60.0 56.4 56.7 57.7 58.5 61.8 55.0 58.4 IntrAgent (ours)75.4 58.2 63.3 65.6 78.5 69.1 70.0 72.5 72.3 67.3 63.3 67.6 Table 23: Accuracy (in percentage) of LUMOS across the three domains inIntraBenchβSERS in chemistry physics (Phys), infectious-disease modeling in public health (PH), and remote sensing in earth science (ES)βunder varying chunk configurations, with GPT-4o, GPT-4.1, and DeepSeek-R1 backbones. LUMOS achieves consistent performance gains as chunk size increases, peaking at 55.8%, 57.7%, and 58.4% in the Phys, PH, and ES domains, respectively. The default 5-chunk setup strikes a balance between accuracy and efficiency across all models. Note. LUMOS with 5 chunks β is the default setting. database. The original PaperQA2 pipeline includes dynamic Paper Search and Citation Traversal mod- ules, which rely on inter-paper citation graphs and large-scale retrieval. Since our corpus consists of a limited number of static papers without citation links, these modules are disabled in our adapta- tion. Instead, we preserve the core reasoning-based retrieval and summarization components of Pa- perQA2, ensuring a faithful yet context-appropriate baseline implementation. Each questionβdocument pair is processed through a streamlined multi-stage pipeline con- sisting of: (1) Document Parsing, where each input PDF is converted to structured text using Grobid; (2) Chunking & Indexing, where parsed text is segmented into semantically coherent units and indexed withFAISS; (3) Dense Retrieval, us- ingintfloat/multilingual-e5-largeto iden- tify the most relevant chunks to the input query; (4) RCS (Contextual Summarization), where GPT-4operforms reasoning-based summarization and relevance scoring to produce high-quality ev- idence summaries; and (5) Answer Generation, where the top-ranked evidence summaries are con- solidated and passed to external LLMs (GPT-4o, GPT-4.1,DeepSeek-R1,Gemini 2.5 Pro,o3, o4-mini,Llama-3.1-70B-Instruct-Turbo) for final response synthesis. All models operate with temperature fixed at 0.0 to ensure deterministic factual outputs. Thedenseretriever (intfloat/multilingual-e5-large)encodes both query and chunk representations under cosine similarity. For each question, the top-k= 30 retrieved chunks are summarized by the RCS stage. The final answer generation step utilizes one of the listed external LLMs, which directly synthesize concise, evidence-grounded answers. The overall model configuration and stage-wise mapping are summarized in Table 24. 38 StageModel / Tool Document Parsing Grobid (structure-aware scientific text parser) Dense Retrieval intfloat/multilingual-e5-large (dual encoder for chunk retrieval) RCS (Contextual Summarization)GPT-4o (reasoning-based summarization and relevance scoring) Final AnswerExternal LLMs: GPT-4o, GPT-4.1, DeepSeek-R1, Gemini 2.5 Pro, o3, o4-mini, Llama-3.1-70B-Instruct-Turbo Table 24: Models used in each stage of the PaperQA2 baseline. F.8 Agentic-Hybrid-Rag This section presents the implementation and evaluation setup of the Agentic-Hybrid-Rag base- line(Nagori et al., 2025), adapted to our exper- imental environment featuring a small, offline corpus consisting of a few domain-specific PDF documents. Since the original Agentic-Hybrid- Rag framework relies on large-scale online bibli- ographic sources (e.g., PubMed, ArXiv, Google Scholar) and a server-based Neo4j knowledge graph, we develop a lightweight variant suitable for a fully local, closed-data scenario. The customized version retains Agentic-Hybrid- Ragβsdual-retrievalarchitectureβcombining graph-basedretrievalandvector-based retrievalβbut replaces all cloud-dependent components with locally executable modules. The system is organized into three main stages: Document Preprocessing,Dual Retrieval, and Final Answer Synthesis. The models and resources used are summarized in Table 25. Each PDF in the corpus is first converted to structured text usingpdfminer.six. Section head- ers and paragraphs are segmented based on typo- graphic cues, and both the extracted text and associ- ated metadata (title, author, year, and keywords) are serialized into JSON format. The resulting dataset thus supports both symbolic (graph) and semantic (vector) indexing without external dependencies. To emulate HybridRAGβs structured reason- ing component without a full Neo4j server, we construct an in-memory knowledge graph using NetworkX. Nodes represent papers, authors, and keywords, while edges encode relations such as written_byandhas_keyword. Given a ques- tion involving bibliographic relations (e.g., βWhich study by Author X discusses. . . ?β), the agent issues rule-based lookups that traverse this lightweight graph and retrieve the corresponding document set. For semantic retrieval, all document sections are embedded using theall-MiniLM-L6-v2sentence transformer and indexed in FAISS. During infer- ence, queries are encoded into the same embedding space, and the top-kpassages are selected based on cosine similarity. To approximate HybridRAGβs hybrid search strategy, BM25 lexical scores are lin- early combined with dense similarity scores before reranking. This design ensures that both surface- level and semantic signals are leveraged despite the small data scale. The retrieved passages from both retrieval modes are merged into a unified context, ordered by rele- vance, and passed to an answer generator. For fair- ness across baselines, we use the same instruction- tuned model (Mistral-7B-Instruct) for all an- swer synthesis tasks. The model is explicitly prompted to reason only over the retrieved evidence to avoid hallucination. The baseline is evaluated on the same set of expert-curated questions used forIntrAgent. For each query, the router decides between graph or vector retrieval depending on the question type. We report accuracy using the LLM-grounded multiple- choice evaluation metric described in Section 5.2. This adaptation allows HybridRAG to operate fully offline while preserving its dual-retrieval reasoning capability, thus serving as a reproducible and inter- pretable baseline for comparison with IntrAgent. F.9 SciMaster We adopt SciMaster (Chai et al., 2025) as the baseline framework for tool-integrated agentic rea- soning. SciMaster enables a reasoning model to dynamically interact with external environments through Python-based code execution, combin- ing tool-augmented reasoning with a multi-stage inference-time workflow involving Solver, Critic, Rewriter, and Selector roles. To adapt SciMaster to our task setting involving domain-specific scientific literature and structured information extraction, we introduce several modifications. First, we perform Domain-Specific Query Adaptation. Each research question from our benchmark is reformulated into SciMasterβs stan- dard input template, consisting of a user query followed by a reasoning block enclosed within 39 StageImplementation / Model Used Graph-based Retrieval NetworkX( Query structured metadata relations) Vector-based Retrieval all-MiniLM-L6-v2+ BM25 + FAISS (Retrieve semantically similar passages using hybrid sparseβdense retrieval) Final AnswerExternal LLMs:GPT-4o,GPT-4.1,DeepSeek-R1,Gemini 2.5 Pro,o3,o4-mini, Llama-3.1-70B-Instruct-Turbo(Synthesize a concise, evidence-grounded response based on all retrieved information) Table 25: Models used in each stage of the Agentic-Hybrid-Rag baseline. <think>and</think>tags. This ensures that the baseline reasoning process remains consistent with the original workflow while aligning the query se- mantics with scientific reading tasks. The same backbone model, DeepSeek-R1-0528, is employed with a temperature of 0.6 and a context length of 64k tokens, identical to the configuration reported in the original paper. We replace the original general-purpose modules web_searchandweb_parsewith a customized local retrieval interface, redirecting all SciMas- ter retrieval requests to our curated corpus of sci- entific papers rather than the open web. This design enables controlled access to factual con- tent and ensures a fair comparison with our task- specific agent operating on the same literature sources. The retrieved text snippets are returned as simulated execution results enclosed within <execution_results>tags, thereby fully preserv- ing the original reasoningβactionβfeedback loop. Following the canonical design, the inference pipeline consists of four sequential stages: (1) the Solver generates multiple candidate reasoning traces; (2) the Critic refines each reasoning trace; (3) the Rewriter synthesizes rewritten versions ref- erencing all prior solutions; and (4) the Selector ranks and outputs the final answer. For efficiency on our dataset, we reduce the number of reasoning iterations per role from five to three while maintain- ing identical temperature and decoding parameters. All stages are executed within a single inference session to ensure consistent context propagation across steps. This adaptation preserves SciMasterβs origi- nal tool-augmented reasoning design while en- suring compatibility with our scientific question- answering setting. The resulting baseline provides a robust reference point for evaluating the perfor- mance gains introduced byIntrAgentβs reasoning- guided section ranking and iterative reading mech- anisms. G Experiments Compute Resources and Model Choices We conducted all experiments using API access to GPT-4o-20240806, GPT-4.1-20250414, and DeepSeek-R1-20250120, o3-20250416, o4-mini- 20250416, gemini-2.5-pro-20250617, and Llama- 3.1-70B-Instruct-Turbo-20250723 β using their de- fault settings. Other experiments compute resources indicated in Table 26. ComponentSpecification GPU ModelNVIDIA RTX 3090 GPU Memory24 GB per GPU CPU Cores22 System Memory64 GB Table 26: Hardware Configuration Used for Training and Inference 40 H Detailed Impact of Mapping Model Across Domains To complement our primary analysis of map- ping model robustness (Section Robustness of IntrAgentto Mapping Model and Input Variabil- ity), we present a comprehensive evaluation across all three scientific domains in IntraBench: SERS in chemistry physics (Phys), infectious-disease modeling in public health (PH), and remote sens- ing in earth science (ES). This analysis inves- tigates how different mapping models perform when paired with short-form answers generated byIntrAgentand vanilla RAG under different backbone LLMs. Specifically, we follow a standard evaluation pro- tocol: short-form answers are first generated by two methods using three backbone LLMsβGPT-4o, GPT-4.1, and DeepSeek-R1. These responses are then mapped to multiple-choice (MCQ) selections using three separate mapping models (GPT-4o, GPT-4.1, and DeepSeek-R1), resulting in 18 unique backbone-mapping combinations. Each configura- tion is evaluated using a consistent prompt template. Accuracy scores for each domain and setting are reported in Table 27. Regardless of the mapping model used, IntrAgentconsistently outperforms the RAG baseline across all domains and backbone config- urations on average. The performance gains re- main substantial, demonstrating the robustness of IntrAgentβs short-form answers across different downstream mapping strategies. Across all nine mappingβbackbone pairs, IntrAgentconsistently exceeds the vanilla RAG baseline: averaged over the three scientific do- mains, its accuracy gains are 4.8%, 10.0%, and 7.7% when GPT-4o is the mapping model (for GPT-4o, GPT-4.1, and DeepSeek-R1 backbones, respectively); 8.0%, 11.0%, and 4.9% when GPT- 4.1 performs the mapping; and 5.6%, 5.0%, and 8.7% when DeepSeek-R1 is used as the mapping model. In addition, we revisited the results of hu- man mapping annotation in collaboration with our physics domain expert. In a prior study, RAG achieved a score of 61 out of 65 when evaluated using GPT-4.1 as the mapping model. Upon closer inspection of the four mismatches, we found that in two cases the modelβs selections were arguably more accurate than those of the human annotator, indicating possible human error when the task is challenging (which causes potential ambiguities in answer interpretation). Nevertheless, even with minor fluctuations in performance across different mapping mod- els, we consistently observe a substantial perfor- mance gap between the best baseline and our ap- proach. This further reinforces our conclusion that IntrAgentachieves state-of-the-art performance in IntraView. I Similarity-Based Section Selection In addition to evaluating baseline approaches, we further investigate how to identify the section of a paper most relevant to a given research question by analyzing section titles. To this end, we conduct an additional experiment comparing two section se- lection strategies. The first is the reasoning-based ranking from the Section Ranking stage developed inIntrAgent. The second is a retrieval-based ap- proach that selects the section whose title exhibits the highest cosine similarity with the research ques- tion, using a vanilla RAG framework. This experiment focuses on the physics dataset, where domain experts have annotated the section titles from which correct answers are expected to originate. We first collect the section titles ranked by the reasoning-based method. For comparison, we replace this with the retrieval-based strategy de- scribed above, computing cosine similarity scores between each question and all section titles, and ranking them in descending order. Both methods are evaluated by comparing their top-3 ranked sec- tion titles against the expert-annotated ground truth. Results are reported in Table 28. Note that in some cases where the correct an- swer is F (none of the above)βindicating that the relevant content does not appear in the paperβwe exclude those questions from the evaluation, as there is no valid section title to compare. As a re- sult, the physics dataset contains 56 questions that are valid for Section Ranking comparison. We observe that our method achieves a top-1 accuracy of 87.5%, a top-2 accuracy of 92.9%, and a top-3 accuracy of 94.6%, in comparison to the RAG-based method, which yields an accuracy of 25.0% at top-1, 32.0% at top-2, and 50.0% at top-3 on the physics dataset. We also evaluate end-to-end task performance using the standard protocol. When replacing our reasoning-based section ranking with the RAG- based reordering strategy, the overall accuracy 41 BackboneGPT-4oGPT-4.1DeepSeek-R1 MappingMethodPhysPHESAvg. PhysPHESAvg. PhysPHESAvg. GPT-4o IntrAgent 73.8 52.7 60.0 62.2 76.9 63.6 68.3 69.6 73.8 67.3 63.3 68.1 RAG63.1 50.9 58.3 57.4 66.2 52.7 60.0 59.6 63.1 58.2 60.0 60.4 GPT-4.1 IntrAgent 75.4 58.2 63.3 65.6 78.5 69.1 70.0 72.5 72.3 67.3 63.3 67.6 RAG60.0 52.7 60.0 57.6 66.2 60.0 58.3 61.5 64.6 63.6 60.0 62.7 DeepSeek-R1 IntrAgent 67.7 58.2 58.3 61.4 70.8 58.2 63.3 64.1 76.9 67.3 55.0 66.4 RAG61.5 50.9 55.0 55.8 66.2 52.7 58.3 59.1 63.1 58.2 51.7 57.7 Table 27: Accuracy (in percentage) of ourIntrAgentand the Vanilla RAG baseline across three scientific domains inIntraBenchβSERS in chemistry physics (Phys), infectious-disease modeling in public health (PH), and remote sensing in earth science (ES). Each row block corresponds to a mapping model (GPT-4o, GPT-4.1, or DeepSeek-R1), and each column block shows performance under a different backbone LLM. Short-form answers from each backbone are mapped to MCQ choices using each mapping model. MethodTop-1 Acc.Top-2 Acc.Top-3 Acc. Similarity-based (RAG)25.0 (14/56)32.0 (18/56)50.0 (28/56) Reasoning-based (IntrAgent)87.5(49/56)92.9(52/56)94.6(53/56) Table 28: Section title ranking accuracy (in percentage) on the physics dataset (Top-1, Top-2, and Top-3). Ground truth is based on expert-annotated sections. drops to 66.2%, compared to 78.5% achieved by IntrAgentusing GPT-4.1 as both the backbone and mapping model. While this substitution leads to a performance decline, the decrease is not sub- stantial. These results highlight the robustness of our iterative reading stage: even when section titles fail to generate an ideal ranking, the model contin- ues to perform strongly by reasoning over multiple retrieved sections. 42