Paper deep dive
CogGen: A Cognitively Inspired Recursive Framework for Deep Research Report Generation
Kuo Tian, Pengfei Sun, Zhen Wu, Junran Ding, Xinyu Dai
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 4/26/2026, 9:15:43 PM
Summary
CogGen is a cognitively inspired recursive framework designed for autonomous deep research report generation using Large Language Models (LLMs). Unlike traditional linear workflows that suffer from error accumulation and rigid structures, CogGen employs a Hierarchical Recursive Architecture consisting of a Macro-Cognitive Loop (for global planning and restructuring) and a Micro-Cognitive Cycle (for parallelized section-level refinement). To handle multimodal content, it introduces Abstract Visual Representation (AVR), which decouples visual intent from rendering to reduce cognitive load. The framework is evaluated using the proposed Cognitive Load Evaluation Framework (CLEF) across five dimensions: Organization, Depth, Relevance, Alignment, and Synergy. Experimental results show CogGen achieves state-of-the-art performance, outperforming open-source systems and matching professional analyst outputs on the Our World in Data (OWID) benchmark.
Entities (10)
Relation Signals (6)
CogGen â evaluatedby â CLEF
confidence 100% ¡ To rigorously evaluate the quality of synthesized reports, we propose the Cognitive Load Evaluation Framework (CLEF).
Writer Agent â generates â Abstract Visual Representation
confidence 100% ¡ The Writer Agent (A w ) is responsible for text composition and the definition of visual intent.
CogGen â incorporates â Abstract Visual Representation
confidence 100% ¡ To extend this recursivity to multimodal content, we introduce Abstract Visual Representation (AVR)
Planner Agent â partof â CogGen
confidence 100% ¡ The process is collaboratively executed by three peer cognitive agents... Planner Agent (A p )
Render Agent â processes â Abstract Visual Representation
confidence 100% ¡ the Render Agent (A render ) translates these semantic intents into executable syntax
CogGen â uses â Hierarchical Recursive Architecture
confidence 100% ¡ Leveraging a Hierarchical Recursive Architecture to simulate cognitive writing, CogGen enables flexible planning...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The autonomous synthesis of deep research reports represents a critical frontier for Large Language Models (LLMs), demanding sophisticated information orchestration and non-linear narrative logic. Current approaches rely on rigid predefined linear workflows, which cause error accumulation, preclude global restructuring from subsequent insights, and ultimately limit in-depth multimodal fusion and report quality. We propose CogGen, a Cognitively inspired recursive framework for deep research report Generation. Leveraging a Hierarchical Recursive Architecture to simulate cognitive writing, CogGen enables flexible planning and global restructuring. To extend this recursivity to multimodal content, we introduce Abstract Visual Representation (AVR): a concise intent-driven language that iteratively refines visual-text layouts without pixel-level regeneration overhead. We further present CLEF, a Cognitive Load Evaluation Framework, and curate a new benchmark from Our World in Data (OWID). Extensive experiments show CogGen achieves state-of-the-art results among open-source systems, generating reports comparable to professional analysts' outputs and surpassing Gemini Deep Research. Our code and dataset are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2604.17072v2
- Canonical: https://arxiv.org/abs/2604.17072v2
Trouble viewing inline? Open PDF directly â
Full Text
104,658 characters extracted from source content.
Expand or collapse full text
CogGen: A Cognitively Inspired Recursive Framework for Deep Research Report Generation Kuo Tian 1,2 , Pengfei Sun 3 , Zhen Wu 1,2 * , Junran Ding 1,2 , Xinyu Dai 1,2 * 1 National Key Laboratory for Novel Software Technology, Nanjing University, China 2 School of Artificial Intelligence, Nanjing University, China 3 Nanjing Haodun Technology Development Co., Ltd. tiank,jrding@smail.nju.edu.cn, chongqingspf@gmail.com,wuz,daixinyu@nju.edu.cn Abstract The autonomous synthesis of deep research reports represents a critical frontier for Large Language Models (LLMs), demanding sophisti- cated information orchestration and non-linear narrative logic. Current approaches rely on rigid predefined linear workflows, which cause error accumulation, preclude global restructur- ing from subsequent insights, and ultimately limit in-depth multimodal fusion and report quality. We propose CogGen, a Cognitively inspired recursive framework for deep research report Generation. Leveraging a Hierarchical Recursive Architecture to simulate cognitive writing, CogGen enables flexible planning and global restructuring. To extend this recursiv- ity to multimodal content, we introduce Ab- stract Visual Representation (AVR): a concise intent-driven language that iteratively refines visual-text layouts without pixel-level regener- ation overhead. We further present CLEF, a Cognitive Load Evaluation Framework, and cu- rate a new benchmark from Our World in Data (OWID). Extensive experiments show CogGen achieves state-of-the-art results among open- source systems, generating reports compara- ble to professional analystsâ outputs and sur- passing Gemini Deep Research. Our code and dataset are available athttps://github.com/ NJUNLP/CogGen. 1 Introduction Driven by advancements in reasoning and tool-use capabilities (OpenAI, 2025d; Anthropic, 2024; Guo et al., 2025), Large Language Models (LLMs) have demonstrated the potential to autonomously syn- thesize structured deep research reports (Zhang et al., 2025; Li et al., 2025b). However, bridging the gap between automated generation and expert- level analytical writing remains a formidable chal- lenge (Zheng et al., 2025; Du et al., 2025). Expert report writing is not a mere assembly of retrieved * Corresponding authors. Human Cognitive ProcessDeep Research Generation PlanWrite Review Multimodal Report Pre-writing Loop Reason Search Update Plan Write Generation Locked Text-only Report Global Restructuring VS Figure 1: Comparison of report writing paradigms. The Human Cognitive Process (left) adopts a recursive âplan- write-reviewâ loop that supports global restructuring throughout the writing process. In contrast, the Deep Research Report Generation (right) relies on a linear workflow, where once the preceding content is gener- ated, it cannot be modified in reverse and limits the generation of subsequent sections. facts; it is a sophisticated cognitive process char- acterized by recursive refinement and the seamless integration of heterogeneous evidence. Existing deep research report generation paradigms primarily fall into two architectural cat- egories: single-agent systems that integrate reason- ing models with complex tool invocation (Google, 2025; OpenAI, 2025c) and multi-agent frameworks that incorporate role-playing coupled with feed- back mechanisms (Shao et al., 2024; Jiang et al., 2024; Wang et al., 2024). Despite being well- designed, both structures typically follow a linear, predefined execution workflow. Once a plan is drafted, the generation follows a forward-only path, making it difficult for existing agent frameworks to perform the âbackward restructuringâ necessary when downstream discoveries invalidate earlier or- ganizational logic (Xu and Peng, 2025). As illus- trated in Figure 1, this linear rigidity stands in stark contrast to the human cognitive writing process, which functions as an inherently non-linear, recur- sive mechanism of exploration. arXiv:2604.17072v2 [cs.MA] 21 Apr 2026 Furthermore, true deep research necessitates the integration of quantitative visual evidence (e.g., charts) to substantiate qualitative claims. However, current multimodal efforts typically generate these elements separately from the text (Shi et al., 2021; Yang et al., 2024). This asynchronous generation creates a superficial relationship between text and image, where a chart might be redundant to the text or lack the specific data granularity mentioned in the narrative. This forces the reader to manually bridge the gap between abstract descriptions and visual data, leading to a fragmented cognitive expe- rience where the visual acts as a mere illustration rather than a synergistic argument. To address these issues, we propose CogGen, a cognitively inspired multi-agent framework emu- lating the recursive nature of expert writing. Draw- ing on the Cognitive Process Theory of Writ- ing (Flower and Hayes, 1981; Hayes, 1996), we introduce a Hierarchical Recursive Architecture. This architecture comprises a Macro-Cognitive Loop for global logic orchestration and a Micro- Cognitive Cycle for autonomous intra-section re- finement. By enabling agents to dynamically pause, review, and restructure the global plan based on emerging information, CogGen transcends the âlin- ear lock-inâ of traditional paradigm, allowing for a fluid and logically coherent narrative evolution. Beyond structural logic, CogGen addresses the multimodal integration gap through the lens of Cog- nitive Offloading (Risko and Gilbert, 2016). Re- search suggests that expert writers often decou- ple high-level content planning from low-level vi- sual rendering to mitigate dual-task interference. Consistent with this behavior, we introduce an Ab- stract Visual Representation (AVR). By abstract- ing verbose visualization specifications into a com- pact intermediate representation, this schema al- lows the agent to treat visual elements as mutable semantic tokens while offloads the final visualiza- tion to specialized rendering agents. This enables the synchronous iteration of narrative and visual plans with minimal cognitive load, ensuring that charts and text achieve a high degree of synergy rather than mere alignment. To rigorously evaluate the quality of synthe- sized reports, we propose the Cognitive Load Evaluation Framework (CLEF). Moving beyond surface-level n-gram metrics, CLEF is grounded in cognitive load theory (Sweller, 1994), assess- ing reports across five dimensions: Organization, Depth, and Relevance, Alignment, Synergy. We benchmark CogGen on a newly curated dataset from Our World in Data (OWID) and the Wild- Seek benchmark. Experimental results demonstrate that CogGen significantly outperforms state-of-the- art open-source frameworks. Notably, CogGen- generated reports achieved parity with human ex- pert benchmarks on OWID and surpassed refer- ences from Gemini Deep Research on WildSeek. Our primary contributions are as follows: ⢠Framework: We propose novel CogGen, a Hierarchical Recursive Framework that opera- tionalizes cognitive writing theories to enable non-linear, global logic restructuring in deep research reports generation. ⢠Mechanism: We introduce an Abstract Visual Representation rooted in cognitive offloading theory, facilitating the deep semantic integra- tion of text and visual evidence. â˘Evaluation: We present CLEF, a cognitive theory-driven evaluation framework, and re- lease a high-quality benchmark based on OWID to facilitate future research in deep research agents. 2 Related Work 2.1 Agentic Report Generation Prior automated report generation primarily re- lied on domain-specific fixed workflows (Wang et al., 2024; Ghafarollahi and Buehler, 2025; Zhang and Eger, 2024; Pichlmair et al., 2024; Huot et al., 2025), whose performance was con- strained by predefined linear processes. Concur- rent works attempt to mitigate this via dynamic retrieval; however, PAGER (Li et al., 2026) targets QA tasks rather than long-form generation, and Mind2Report (Cheng et al., 2026) retains a unidi- rectional serial workflow lacking global restructur- ing. To address complex tasks, frameworks like WriteHere (Xiong et al., 2025) and ReCode (Yu et al., 2026) introduce recursive decomposition. Yet, they remain essentially forward-generation methods unable to retroactively resolve structural disruptions. Similarly, while ARCS (Bhattarai et al., 2025) utilizes execution-repair loops, its global granularity scales poorly to comprehensive reports. Other studies enhance planning via role- playing (Shao et al., 2024; Jiang et al., 2024), fail- ing to address the disconnect between writing and planning. Furthermore, despite recent advances Query Writer Agent (í´ í¤ ) (See Box B) Search (íž) Replan (í Ⲡí ) Write (íś í ) Render Agent (í´ íííííí ) Multimodal Deep Research Report Macro Iteration (íĄâíĄ+1) Iterative Draft Generation (Macro-Cognitive Loop) Planner Reviewer Writer Global Outline (í íĄ ) Draft (íś íĄ ) Feedback (â) Writer Agent (í¨ í ) (Micro-Cognitive Cycle) Thread s (Section s) Visual Render Approved Draft B C A Reviewer Agent (í´ í ) Planner Agent (í´ í ) Global Search (íž) Structure Plan (í íĄ ) Merge Draft íś íĄ Visual View Figure 2: Overview of the CogGen framework. Components marked with an eye icon indicate operations strictly monitored by the Reviewer Agent (A r ) to enable feedback-driven iteration. (A) Macro-Cognitive Loop: A global iterative process consisting of three phases. The Planner Agent (A p ) generates the outline (O (t) ), the Writer Agent (A w ) produces the draft (C (t) ), and the Reviewer Agent (A r ) evaluates the complete draft to generate feedback (â) for the next iteration. (B) Micro-Cognitive Cycle: Within the Writer Agent (A w ), multiple threads execute monitored âSearchâReplanâWriteâ cycles to generate section drafts (C s ), which are finally merged into the draft (C t ). (C) Visual Rendering: In the Execution phase, the Render Agent (A render ) translates the approved draft into a visual view, operating under the Reviewerâs supervision to ensure alignment with the visual specifications. in verification-centric evaluations like DEER (Han et al., 2025), even state-of-the-art commercial mod- els (e.g., OpenAI (OpenAI, 2025c) and Gemini Deep Research (Google, 2025)) remain limited by fixed frameworks during their writing execution stage. In contrast, CogGen proposes a recursive outline modification mechanism (Global Restruc- turing) to iteratively refine both historical and fu- ture content contextually. 2.2 Multimodal Report Generation Early multimodal report generation primarily re- lied on domain-specific frameworks (Shi et al., 2021; Yang et al., 2024), adopting a sequential slot- filling strategy to generate text and visuals indepen- dently. Recent works such as Multimodal DeepRe- searcher (Yang et al., 2025a) enabled open-domain multimodal generation by introducing visual de- scription languages (Satyanarayan et al., 2017) and embedding chart generation into linear workflows. However, they are essentially loose combinations of text and visual generation without in-depth col- laborative optimization. In contrast, CogGen in- troduces the Abstract Visual Representation and shifts the objective from visual fidelity to the char- acterization of visual semantic intent, achieving semantic-level collaborative planning and iterative optimization of both textual and visual content. 3 Methodology 3.1 Framework Overview To overcome the linear constraints discussed in Sec- tion 1, CogGen implements Hierarchical Recursive Architecture (Figure 2). Instead of a static chain, this design treats the generation plan as a mutable object, enabling dynamic, non-linear transitions across planning, writing, and reviewing phases. Formally, we model report generation as a map- ping from a user queryQto a multimodal deep research reportR, denoted asR = CogGen(Q). The process is collaboratively executed by three peer cognitive agents (Figure 2A): â˘Planner Agent (A p ): Responsible for infor- mation retrieval and structural planning. Its function is formalized as a mappingO,K = A p (Q,H), whereHrepresents the interaction history and feedback state.Qis the user query, Ois the writing outline, andKis the knowl- edge base formed by information retrieved during outline generation. â˘Writer Agent (A w ): Responsible for text com- position and the definition of visual intent. Its function is formalized asC = A w (O,K), whereCrepresents the draft with the abstract vision representations (AVRs) generated by the writing agent. ⢠Reviewer Agent (A r ): An integrated eval- uation engine with dual functions of real- time monitoring and post-hoc assessment. By outputting feedback signalsâ, this agent achieves two core objectives: ensuring the generation process adheres to preset con- straints under monitoring mode, and optimiz- ing content quality under reviewing mode. Unlike traditional linear chain structures (Shao et al., 2024; Yang et al., 2025b), this collabora- tive agent triad supports recursive operations at both the macro (global report) and micro (local sec- tion) granularities, as illustrated in parts A and B of Figure 2, ensuring generation quality through immediate review mechanisms. 3.2 Macro-Cognitive Loop The core engine of CogGen is designed to enable Global Restructuring. To address the rigidity of linear workflows, where the generated preceding content cannot be reconstructed in reverse (Xu and Peng, 2025), CogGen utilizes a Macro-Cognitive Loop to implement recursive optimization. This mechanism empowers the system to per- form backward restructuring: it allows agents to retroactively refine the global outline (O) and previ- ously generated drafts based on downstream discov- eries. This ensures that the final report maintains global logical coherence rather than being a linear accumulation of sub-tasks. In the loop shown in Figure 2, t represents the iteration round. 3.2.1 Iterative Global Planning The process begins with macro planning. First, the Planner Agent (A p ) performs a breadth-first retrieval to construct the initial knowledge base Kand a report blueprint, denoted formally as the outlineO (0) . This corresponds to the initial state where history is empty (H =â ): O (0) ,K = A p (Q,â )(1) To support parallel generation (Section 3.3),K adopts a hierarchical architecture: a shared global Structure of Abstract Visual Representation (P vis ) [DATA_VISUALIZATION] Title: Adoption of Key AI Technologies in Michelin... Chart_Type: Bar Chart X_Axis: Types of AI Technology (Chatbots, Robotics... Y_Axis: Estimated Adoption Level in Restaurants... Data_Source: <ref:1003> Purpose: To visually compare the adoption rates... [/DATA_VISUALIZATION] Table 1: An instantiation of the Abstract Visual Repre- sentation (AVR). The Writer generates this structured semantic representation instead of executable code, de- coupling reasoning from rendering. snapshot provides common context to all gener- ation threads, while section-specific evidence re- trieved during micro-cycles is maintained in thread- local caches. This design prevents irrelevant noise from propagating across unrelated chapters while ensuring each thread retains the targeted evidence required for deep synthesis. A formal specification of this protocol is provided in Appendix B.1. In subsequent rounds (t > 0),A p refines the structure based on the feedback signalâ (t) de- rived from the previous draftC (t) . This constitutes the âMacro-Cognitive Loopâ (Figure 2A), enabling retroactive adjustments to global logic: O (t+1) = A p (Q,O (t) , â (t) | K)(2) This recursive update ensures that the narrative structure and visual planning co-evolve, preventing the logical inconsistencies typical of static planning approaches. 3.2.2 Parallel Multimodal Content Writing To improve report synthesis efficiency, CogGen generates multiple sections in parallel (details are shown in Section 3.3). Specifically, the Writer AgentA w generates a unified draftC (t) based on the global outline. To ensure parallel consistency, the generation of each section strictly follows the constraints of the global outlineO (t) : C (t) =A w (o s ,O (t) ,K)|âo s âO (t) (3) By using the global structureO (t) as a constraint, all parallel generation threads maintain consistency with the overall logic of the report. The draft C (t) contains both textual content and AVRs (P vis ). These vision representations carry complete visual- ization intents (shown in Table 1) but use a highly structured description to reduce cognitive load. 3.2.3 Global Review The Reviewer AgentA r conducts a comprehensive evaluation of the current draftC (t) and outputs a feedback signalâ (t) . This signal contains opti- mization suggestions for the current outline based on the newly generated draft. The feedback signal â (t) serves as the input for the next round of plan- ning, thereby driving the co-evolution of text and visual content through the recursive loop. To enforce stability, CogGen incorporates a strict monotonic improvement constraint. Rather than re- lying on open-ended refinement, the system accepts a global update only when the Reviewer Agent validates a distinct increase in report quality. By rejecting changes that fail to meet this evaluation threshold, the architecture is designed to suppress infinite oscillation and drive the draft towards a local optimum relative to the reviewerâs criteria. Appendix A provides a theoretical analysis of the convergence properties of this mechanism, model- ing CogGen as a bounded state-space search with empirically validated stability. 3.3 Micro-Cognitive Cycle While the macro mechanism maintains global co- herence, the detailed content generation is handled via parallelized micro-cycles. As illustrated in Part B of Figure 2, the Writer Agent does not generate linearly; instead, it orchestrates multiple indepen- dent threads in parallel, recursively invoking the capabilities of the Planner and Reviewer Agents. Recursive Execution Flow. Consistent with the workflow depicted in Figure 2, each section generation thread (Thread s ) executes a recursive âSearchâReplanâWriteâ process: ⢠Search and Replan: The thread temporarily re-engages the Planner Agent to perform tar- geted retrieval and, if necessary, adaptively adjusts the sectionâs internal outline based on retrieved evidence. â˘Write: The Writer Agent then composes the section text based on the retrieved evidence and refined outline. â˘Review: The search, replan, and write pro- cesses are continuously monitored by the Re- viewer Agent. Any intermediate state or final content that deviates from the requirements triggers an immediate correction loop, ensur- ing that errors are caught and resolved before propagating to the next stage. Parallelism and Deferred Update. Integrating retroactive revision into a serial workflow intro- duces critical stability issue we term Contextual Os- cillation: correcting an upstream section (e.g., Sec 1) to align with a downstream discovery (e.g., Sec 5) invalidates the intermediate context. Without a global perspective, the model performs myopic correctionsâfixing Sec 1 creates new inconsisten- cies with Sec 5, triggering a recursive modification loop between chapters (Huang et al., 2024). Since the draft is incomplete during this serial process, the agent lacks the holistic view required to re- solve these cross-section conflicts, leading to non- convergence. To break the issue from recursive loops inherent in serial revision, CogGen employs a parallel ar- chitecture with a Deferred Update Policy: parallel micro-cycles operate as read-only observers of the global outlineO (t) , with section-specific retrieval confined to thread-local caches. Cross-section con- flicts are not resolved locally but deferred to the Re- viewer AgentA r , which serves as the sole arbitra- tor during macro-cycle transitions (Appendix B.1). Under this policy,A r aggregates all cross-section conflicts into a global feedback signal â (t) . â (t) â A r (C (t) ,O (t) )(4) This signal provides high-level guidance for the subsequent replanning phase (O (t+1) ). By resolv- ing conflicts at the global outline level rather than the local text level, CogGen ensures that structural adjustments are coherently propagated across all dependent chapters. A theoretical analysis of con- vergence properties is provided in Appendix A, with empirical validation in Appendix B. 3.4 Visual Rendering Engine To efficiently handle multimodal fusion, we op- erationalize the Cognitive Offloading strategy pro- posed in Section 1. Instead of disrupting the reason- ing flow with complex code generation (Sweller, 1994), the Writer Agent (A w ) employs an Abstract Visual Representation mechanism. It focuses solely on the visual intent (P vis )âdescribing data points and chart types without implementation details (as shown in Table 1, detailed in appendix F). This design contrasts with the Formal Descrip- tion of Visualization (FDV) adopted by prior work (Yang et al., 2025b): while FDV requires the Writer to simultaneously specify visual styling, layout, and data, AVR captures only semantic in- DimensionEvaluation Focus D1: OrganizationHierarchical structure and navigation D2: DepthCausal explanations and schema con- struction D3: RelevanceAppropriate complexity and coherence D4: Alignment Spatial/semantic integration of visuals and text D5: SynergyInformation complementarity beyond text Table 2: Overview of CLEFâs five evaluation dimen- sions grounded in Cognitive Load Theory. tent (what to show and why), offloading visual de- sign decisions to a dedicated Render Agent. This separation of concerns frees the Writerâs cognitive resources for narrative reasoning and provides a natural insertion point for post-rendering data veri- fication. A quantitative comparison is presented in Section 5.4. Subsequently, the Renderer Agent (A render ) acts as a code interpreter, translating these semantic in- tents into executable syntax (P syn ) using libraries such as ECharts (Li et al., 2018) or Mermaid (Svei- dqvist and Team, 2014). This generation process includes a syntax validation check to ensure exe- cutability before rendering the final style-consistent visual assets (V) in a headless browser.The pipeline is formalized as: P syn = A render (P vis ) V = Browser(P syn ) (5) This two-stage rendering scheme reduces the cognitive load during the writing and planning phases by decoupling the visual planning and gen- eration stage from the rendering stage. 4 Experimental Setup In this section, we detail the experimental configu- ration used to evaluate CogGenâs performance. We first introduce the two datasets used for evaluat- ing report generation capabilities, then define the baseline models used for comparison, and finally elaborate on our proposed evaluation metrics based on cognitive load theory (Sweller, 1994). 4.1 Datasets To comprehensively evaluate the modelâs capabil- ity in generating high-quality deep research reports, we employ a hybrid evaluation strategy combin- ing a self-constructed dataset with an established benchmark. Given the scarcity of existing datasets containing professional-grade reports with rich data visualizations, we curated the OWID dataset to serve as a gold standard for complex multimodal generation. Complementarily, we adopt WildSeek, a standard dataset from prior work (Jiang et al., 2024), to assess the modelâs robustness in handling diverse user intents within open-domain scenarios. OWID.This dataset contains 40 research reports collected from the Our World in Data (OWID) web- site. Written by professional analysts, these reports feature substantial data density and logical depth, and include rich data visualizations. Detailed pro- cedures for dataset construction and preprocess- ing are provided in Appendix G. We use these reports as the Human Gold-Standard to evaluate the modelâs ability to generate comprehensive and high-quality multimodal content. WildSeek. WildSeek (Jiang et al., 2024) was originally a standard dataset for evaluating pure text report generation. To adapt to the objectives of this study, we manually selected 20 queries with clear multimodal generation tendencies (e.g., questions requiring trend comparison or distribution display) to test the robustness of the model in generating illustrated reports in open-domain scenarios. 4.2 Baselines We benchmark CogGen against a comprehensive set of baselines representing distinct generation paradigms: (1) STORM (Shao et al., 2024) and Co- STORM (Jiang et al., 2024), the standard baselines for multi-perspective QA and collaborative writ- ing; (2) WriteHere (Xiong et al., 2025), the current state-of-the-art open-source model; and (3) Multi- modal DeepResearcher (Yang et al., 2025a), which represents linear multimodal generation workflows. Reference Standards. For the OWID dataset, human-authored reports serve as the gold stan- dard. For the WildSeek dataset, which lacks hu- man ground truth, we adhere to established proto- cols (Du et al., 2025) by employing outputs from Gemini Deep Research (Google, 2025) as a com- mercial reference anchor for scoring. 4.3 Metrics: Cognitive Load Evaluation Existing evaluation metrics present significant lim- itations when applied to multimodal deep research reports. Mechanical metrics (Papineni et al., 2002; Lin, 2004) focus on textual n-gram overlap, failing to capture the quality of text and visual elements from semantics. Similarly, while standard LLM- as-a-Judge approaches (Zheng et al., 2023) assess ModelOrganizationDepthRelevanceAlignmentSynergyAvg. Score Dataset I: OWID (High-Density Multimodal Reports) Human Gold-Standard (Ref)0.49860.50000.50000.50000.50000.4997 STORM0.42530.44430.39860.16750.16670.3205 Co-STORM 0.41320.42610.42810.17940.16670.3227 Multimodal DeepResearcher0.37680.42930.35080.18190.17000.3018 WriteHere0.49120.55030.49360.38460.33120.4502 CogGen (Ours)0.49720.58130.50420.48060.43260.4992 Dataset I: WildSeek (Text-Centric Complex Queries) Gemini Deep Research (Ref)0.50000.50000.50000.50000.50000.5000 STORM0.43750.40970.44720.19030.19080.3351 Co-STORM0.39930.36950.42700.19430.18340.3147 Multimodal DeepResearcher0.38190.37400.36950.20760.21830.3103 WriteHere0.52430.49310.52710.47380.44970.4936 CogGen (Ours)0.53890.50000.53340.55440.54370.5341 Table 3: Main Results on Multimodal Report Generation. Scores represent the Relative Advantage Score (R) based on pairwise comparison against the Reference (Ref). A score of 0.5000 indicates parity with the reference; values > 0.5 indicate the model outperforms the reference. CogGen achieves comparable performance to Human Experts in overall quality (Avg. Score) on the data-intensive OWID dataset, driven by superior Depth and Relevance, and outperforms Gemini Deep Research on the text-centric WildSeek dataset. The best results are highlighted in bold, and the second-best are underlined. general semantic quality, they lack a theoretical grounding to evaluate the cognitive synergy be- tween modalities. Specifically, whether visual aids reduce the readerâs mental effort. To bridge these gaps, we propose the Cognitive Load Evaluation Framework (CLEF), grounded in Cognitive Load Theory (Sweller, 1994) and Mayerâs Cognitive The- ory of Multimedia Learning (Mayer, 2005). CLEF operationalizes 11 of Mayerâs 14 multi- media principles into five orthogonal evaluation di- mensions. Table 2 provides an overview of the five dimensions. These dimensions are organized into two categories: Control Dimensions (D1-D3) ensur- ing general content quality, and Core Dimensions (D4-D5) focusing on multimodal integration qual- ity. Three CTML principles (Modality, Temporal Contiguity, Voice) are excluded as they specifically address dynamic multimedia and are not applicable to static text-visual reports. Notably, our evalua- tion framework explicitly classifies tables as visual modalities. This decision is grounded in Cognitive Load Theory, which posits that tabular organiza- tionâlike graphical elementsâsignificantly miti- gates cognitive load. While CLEF operationalizes established cognitive principles into measurable dimensions rather than directly measuring reader behavior (e.g., subjective workload), its validity is supported by high consistency with human ex- pert judgments (Section 5.3) and robustness across multiple evaluation models (Appendix C). Following recent best practices (Du et al., 2025; Krumdick et al., 2025), we employ pairwise com- parison using GPT-5 (OpenAI, 2025b) as the evalu- ator. For each dimension, we calculate the Relative Advantage Score (R â [0, 1]), whereR > 0.5 indicates the model outperforms the baseline in enhancing understanding or reducing cognitive bur- den. Complete theoretical foundations, detailed dimension definitions, scoring mechanisms, and validation results are provided in Appendix D. 4.4 Implementation Details CogGen is implemented using a multi-agent ar- chitecture.The search tool utilizes GPT-4.1- Mini (OpenAI, 2025a) for cost-effective query ex- pansion, while the Planner, Writer, Reviewer and Render Agents utilize GPT-4.1 to ensure reasoning depth. To balance generation diversity and stability, we set the temperature to 0.5 for all agents. The external retrieval tool is the Tavily Search (Tavily, 2025). Notably, for fair comparison, the backbone LLM of baselines were unified to GPT-4.1, and the retrieval tool was unified to Tavily Search. 5 Results and Analysis 5.1 Main Experimental Results Table 3 presents the Relative Advantage Scores cal- culated based on the CLEF evaluation metrics. The experimental results show that CogGen exhibits Method Variants Core MechanismsEvaluation Metrics (Relative to Full Model) Avg. Score Cog. Loop Native MMOrganization Depth Relevance Alignment Synergy GPT-4.1 (W/Search)Ă0.47220.40800.48750.35190.34000.4119 CogGen-no-reviewĂâ0.46110.45480.48890.50020.43560.4681 CogGen-TwoStageâĂ0.48930.51670.49440.46270.48900.4904 CogGenâ0.49860.50000.49860.50000.50000.4994 Table 4: Ablation Study Results on OWID dataset. Cog. Loop, Cognitive Loop denotes the reviewer-driven dynamic modification, and Native M, Native Multimodality refers to the synchronous text-image collaborative planning (via AVR). Scores denote Relative Advantage using CogGen as the reference. significant advantages in tests on both the OWID and WildSeek datasets. On the OWID dataset, CogGen demonstrates strong generation capabilities, achieving evalua- tion scores approaching the Human Gold-Standard while significantly outperforming baseline models such as Multimodal Deep Researcher and Write- Here. Regarding multimodal alignment, although CogGen slightly trails human experts, it secures su- perior synergy scores compared to all baselines. This advantage is driven by the AVR strategy, which enables iterative coordination between tex- tual and visual planning. Notably, CogGen sur- passes human references in Depth. We attribute this to that CogGen explicitly provides broader causal context and background information, resulting in higher informational density. Experiments on the WildSeek dataset further verify the generalization ability of CogGen. With Gemini Deep Research as the reference benchmark, CogGen achieves the highest scores in all five evalu- ation dimensions. Although Gemini reports narrow the score gap in the multimodal dimension through rich tabular content, their shortcoming of lacking adaptive narrative ability is still obvious. In con- trast, baseline models such as WriteHere adopt a recursive decomposition strategy but lack a retroac- tive rewriting mechanism, leading to fragmented report structures. In comparison, CogGen relies on a hierarchical recursive mechanism to dynamically adjust the outline, ultimately achieving comprehen- sive leadership in all five dimensions. 5.2 Ablation Study Table 4 details the comparative performance of CogGen against a Retrieval-Augmented GPT-4.1 baseline and its own ablation variants. In direct comparison, the full CogGen framework demon- strates a comprehensive advantage over the GPT- 4.1 baseline across all evaluation metrics. Most notably, we observe significant gains in Depth and Synergy, validating that our recursive architecture outperforms standard linear RAG workflows in han- dling complex, multimodal synthesis tasks. To isolate the specific contributions of our archi- tectural innovations, we conducted ablation studies focusing on two critical mechanisms: (1) Cogni- tive Loop: reviewer-driven recursive modification. (2) Native Multimodality: text-image collaborative planning via the AVR strategy. We implemented two variants to verify whether these mechanisms are essential for enhancing content quality and en- suring high-quality visual integration. CogGen-no-review: This variant removes the recursive modification mechanism for the outline, retaining only the iterative retrieval and parallel section writing functions. Experimental results in- dicate that after removing the recursive modifica- tion mechanism, the modelâs scores in the three metrics of Organization, Depth, Synergy all show a significant decline; while the scores of Alignment and Relevance remain basically stable. This result shows that the core role of the review module is to improve the global content organization ability and analysis performance of the report, while the writing quality of local content mainly depends on the inherent capabilities of the model. CogGen-TwoStage: This variant removes the AVR-based image-text coordination from the plan- ning and generation phases. It employs a âtext-first, image-laterâ strategy, where the model first gen- erates a plain text report before embedding AVR- driven visualizations for final rendering. Experi- mental data shows that this two-stage generation pipeline results in the most significant drop in the Alignment metric, because the post-inserted im- ages struggle to achieve coherent semantic align- ment with the textual content. Synergy has a slight decline, as the text-derived visualizations still effec- tively reduce cognitive load despite lacking explicit alignment. Notably, the Content Depth of this two- stage variant even surpasses that of the full model. This result aligns with our hypothesis: decoupling visual constraints reduces the cognitive load during text generation, enhancing the depth of analysis. 5.3 Human Evaluation We further conducted a blinded head-to-head hu- man evaluation of CogGen against the baseline model Multimodal DeepResearcher (MMDR) and the proprietary closed-source model Gemini Deep Research on the WildSeek dataset, with assess- ments carried out across four dimensions: Depth, Alignment, Synergy, and Overall Quality. CogGen achieved a dominant 90% win rate over Multimodal DeepResearcher in terms of Overall Quality. Notably, against Gemini Deep Research, CogGen maintained a significant edge in both Over- all Quality (75% win rate) and Multimodal Synergy (80% win rate); additionally, despite being built on a weaker base model, CogGen attained comparable reasoning depth to Gemini (50% win rate). Human evaluation results and automatic evaluation results in Table 3 consistently validate the effectiveness of the proposed hierarchical recursive framework CogGen (see Appendix C.2 for details). Bootstrap significance analysis (B=10,000) further confirms that CogGen is the only system with no significant difference from the human reference level (p=0.88; Appendix C.4). Additionally, factuality evaluations confirm CogGenâs reliability, achieving the high- est citation precision and human-verified supported rate among all compared systems (Appendix E). 5.4 Efficacy of AVR To validate the Abstract Visual Representation (AVR), we compare it with the Formal Descrip- tion of Visualization (FDV) used in MMDR. By capturing only semantic intent rather than full vi- sual specification, AVR significantly reduces the cognitive burden on the Writer, freeing it from vi- sual design dutiesâa factor we argue mitigates the Dual-Task Interference reflected in MMDRâs lower scores across all dimensions in Table 3. The ablation results in Section 5.2 corroborate this hy- pothesis. Beyond reducing cognitive load, AVRâs decou- pled architecture directly addresses the critical is- sue of chart data hallucination. As shown in Ta- ble 5, while AVR without verification exhibits hallucination rates comparable to FDV (67% vs. 60%), its lightweight format provides a natural insertion point for a Post-Rendering Audit. By cross-checking the rendered data points against the ConfigurationHalluc. No Halluc. FDV (MMDR)60%40% AVR w/o verification67%33% AVR + verification28%72% Table 5: Chart data hallucination rates across visual- ization strategies. AVR without verification has com- parable hallucination rates to FDV, but the decoupled architecture enables a Post-Rendering Audit that sub- stantially reduces hallucination. knowledge base, this verification-in-the-loop mech- anism substantially reduces the final hallucination rate to 28%. This demonstrates that AVR is a struc- tural enabler for reliable multimodal generation. For detailed token-level cognitive load analysis, see Appendix F.2. 6 Conclusion This paper presents CogGen, a cognitively inspired framework that overcomes the linear execution con- straints of current deep research agents. By inte- grating a Hierarchical Recursive Architecture with a Parameterized Placeholder Mechanism, CogGen enables non-linear logic restructuring and synergis- tic multimodal integration. Our evaluation via the CLEF framework and OWID benchmark demon- strates that CogGen achieves performance com- parable to human experts in analytical depth and multimodal synergy. These findings validate the ef- ficacy of cognitive architectures in evolving LLMs from linear executors into autonomous, recursive researchers. Acknowledgments We thank the anonymous reviewers and the area chair for their constructive feedback, which signifi- cantly improved this paper. This work is supported by the NSFC (No. 62376120, 62576163). Limitations While CogGen introduces parallelized generation to improve efficiency, the introduced recursive mechanisms incur additional computational over- head. Furthermore, constrained by current gener- ation and rendering bottlenecks, there remains a quality gap between our automated charts and those curated by human experts. Additionally, the current rendering pipeline deliberately restricts the Render Agent to high-level declarative libraries (ECharts and Mermaid) to ensure stability; this design choice limits the expressiveness for highly customized sci- entific visualizations achievable through imperative programming. Ethical considerations We prioritize ethical responsibility throughout the frameworkâs development. Regarding informa- tion veracity, we acknowledge that despite veri- fication mechanisms, LLMs may produce hallu- cinations; thus, generated reports should serve as references requiring human oversight rather than absolute truths, and we caution against potential misuse for disinformation. In terms of data pri- vacy, we rigorously filtered our dataset to exclude Personally Identifiable Information (PII) and uti- lized commercial APIs in compliance with usage policies. Finally, our human evaluation involved graduate student volunteers who participated with full knowledge of the studyâs purpose and with- out financial compensation, ensuring adherence to ethical standards for user studies. References Anthropic. 2024. Claude 3.5 Sonnet. Technical report, Anthropic.https://w.anthropic.com/news/ claude-3-5-sonnet. Manish Bhattarai, Miguel Cordova, Minh Vu, Javier Santos, Ismael Boureima, and Dan OâMalley. 2025. ARCS: Agentic Retrieval-Augmented Code Synthesis with Iterative Refinement. Preprint, arXiv:2504.20434. Mingyue Cheng, Daoyu Wang, Qi Liu, Shuo Yu, Xiaoyu Tao, Yuqian Wang, Chengzhong Chu, Yu Duan, Mingkang Long, and Enhong Chen. 2026. Mind2Report: A Cognitive Deep Research Agent for Expert-Level Commercial Report Synthe- sis. Preprint, arXiv:2601.04879. Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, and Zhendong Mao. 2025. DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents. Preprint, arXiv:2506.11763. Linda Flower and John R. Hayes. 1981. A Cognitive Process Theory of Writing. College Composition and Communication, 32(4):365â387. Alireza Ghafarollahi and Markus J. Buehler. 2025. Sci- Agents: Automating Scientific Discovery Through Bioinspired Multi-Agent Intelligent Graph Reason- ing. Advanced Materials, 37(22):2413523. Google. 2025. Gemini deep research â your personal research assistant. Technical report, Google.https: //gemini.google/overview/deep-research/. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhu- oshu Li, Ziyi Gao, Aixin Liu, and 180 others. 2025. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. Preprint, arXiv:2501.12948. Janghoon Han, Heegyu Kim, Changho Lee, Dahm Lee, Min Hyung Park, Hosung Song, Stanley Jungkyu Choi, Moontae Lee, and Honglak Lee. 2025. DEER: A Comprehensive and Reliable Bench- mark for Deep-Research Expert Reports. Preprint, arXiv:2512.17776. John R. Hayes. 1996. A new framework for understand- ing cognition and affect in writing. In The Science of Writing: Theories, Methods, Individual Differences, and Applications, pages 1â27. Lawrence Erlbaum Associates, Inc, Hillsdale, NJ, US. JieHuang,XinyunChen,SwaroopMishra, Huaixiu Steven Zheng, Adams Wei Yu, Xiny- ing Song, and Denny Zhou. 2024. Large Language Models Cannot Self-Correct Reasoning Yet. Preprint, arXiv:2310.01798. Fantine Huot, Reinald Kim Amplayo, Jennimaria Palo- maki, Alice Shoshana Jakobovits, Elizabeth Clark, and Mirella Lapata. 2025. Agentsâ Room: Narra- tive Generation through Multi-step Collaboration. Preprint, arXiv:2410.02603. Yucheng Jiang, Yijia Shao, Dekun Ma, Sina J. Sem- nani, and Monica S. Lam. 2024. Into the Unknown Unknowns: Engaged Human Learning through Par- ticipation in Language Model Agent Conversations. Preprint, arXiv:2408.15232. Michael Krumdick, Charles Lovering, Varshini Reddy, Seth Ebner, and Chris Tanner. 2025. No free labels: Limitations of llm-as-a-judge without human ground- ing. arXiv preprint arXiv:2503.05061. Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, Kuan Li, Liangcai Su, Litu Ou, Liwen Zhang, Pengjun Xie, Rui Ye, Wenbiao Yin, Xinmiao Yu, Xinyu Wang, Xixi Wu, and 37 oth- ers. 2025a. Tongyi deepresearch technical report. Preprint, arXiv:2510.24701. Deqing Li, Honghui Mei, Yi Shen, Shuang Su, Wenli Zhang, Junting Wang, Ming Zu, and Wei Chen. 2018. ECharts: A declarative framework for rapid construc- tion of web-based visualization. Visual Informatics, 2(2):136â146. Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. 2025b. Search-o1: Agentic Search- Enhanced Large Reasoning Models. Preprint, arXiv:2501.05366. Xinze Li, Zhenghao Liu, Haidong Xin, Yukun Yan, Shuo Wang, Zheni Zeng, Sen Mei, Ge Yu, and Maosong Sun. 2026. Structured Knowledge Rep- resentation through Contextual Pages for Retrieval- Augmented Generation. Preprint, arXiv:2601.09402. Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74â81. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paran- jape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language mod- els use long contexts. Transactions of the Association for Computational Linguistics, 12:157â173. Richard E Mayer. 2005. Cognitive theory of multimedia learning. The Cambridge handbook of multimedia learning, 41(1):31â48. OpenAI. 2025a. GPT-4.1. Technical report, OpenAI. https://openai.com/index/gpt-4-1/. OpenAI. 2025b.GPT-5.https://openai.com/ index/introducing-gpt-5/. OpenAI. 2025c.OpenAI deep research.Techni- cal report, OpenAI.https://openai.com/index/ introducing-deep-research/. OpenAI. 2025d. OpenAI O1 system card. Techni- cal report, OpenAI.https://openai.com/index/ openai-o1-system-card/. Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic evalu- ation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computa- tional Linguistics, pages 311â318. Martin Pichlmair, Riddhi Raj, and Charlene Putney. 2024. Drama Engine: A Framework for Narrative Agents. Preprint, arXiv:2408.11574. Evan F. Risko and Sam J. Gilbert. 2016. Cognitive Offloading. Trends in Cognitive Sciences, 20(9):676â 688. Arvind Satyanarayan, Dominik Moritz, Kanit Wong- suphasawat, and Jeffrey Heer. 2017. Vega-lite: A grammar of interactive graphics. IEEE Trans. Visual- ization & Comp. Graphics (Proc. InfoVis). 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. Danqing Shi, Xinyue Xu, Fuling Sun, Yang Shi, and Nan Cao. 2021. Calliope: Automatic Visual Data Story Generation from a Spreadsheet. IEEE Trans- actions on Visualization and Computer Graphics, 27(2):453â463. Knut Sveidqvist and Mermaid Development Team. 2014.Mermaid: Generation of diagrams and flowcharts from text. Software, MIT License. John Sweller. 1994. Cognitive load theory, learning difficulty, and instructional design. Learning and instruction, 4(4):295â312. Tavily. 2025.Tavily search api.https://docs. tavily.com/documentation. Yidong Wang, Qi Guo, Wenjin Yao, Hongbo Zhang, Xin Zhang, Zhen Wu, Meishan Zhang, Xinyu Dai, Min Zhang, Qingsong Wen, Wei Ye, Shikun Zhang, and Yue Zhang. 2024. AutoSurvey: Large Language Models Can Automatically Write Surveys. Preprint, arXiv:2406.10252. Ruibin Xiong, Yimeng Chen, Dmitrii Khizbullin, Mingchen Zhuge, and JĂźrgen Schmidhuber. 2025. Beyond Outlining: Heterogeneous Recursive Plan- ning for Adaptive Long-form Writing with Language Models. Preprint, arXiv:2503.08275. Renjun Xu and Jingwen Peng. 2025. A Comprehensive Survey of Deep Research: Systems, Methodologies, and Applications. Preprint, arXiv:2506.12594. Hongyang Yang, Boyu Zhang, Neng Wang, Cheng Guo, Xiaoli Zhang, Likun Lin, Junlin Wang, Tianyu Zhou, Mao Guan, Runjia Zhang, and Christina Dan Wang. 2024. FinRobot: An Open-Source AI Agent Plat- form for Financial Applications using Large Lan- guage Models. Preprint, arXiv:2405.14767. Zhaorui Yang, Bo Pan, Han Wang, Yiyao Wang, Xingyu Liu, Minfeng Zhu, Bo Zhang, and Wei Chen. 2025a. Multimodal DeepResearcher: Generating Text-Chart Interleaved Reports From Scratch with Agentic Framework. Preprint, arXiv:2506.02454. Zhongyu Yang, Jun Chen, Dannong Xu, Junjie Fei, Xiaoqian Shen, Liangbing Zhao, Chun-Mei Feng, and Mohamed Elhoseiny. 2025b. Wikiautogen: To- wards multi-modal wikipedia-style article generation. Preprint, arXiv:2503.19065. Zhaoyang Yu, Jiayi Zhang, Huixue Su, Yufan Zhao, Yifan Wu, Mingyi Deng, Jinyu Xiang, Yizhang Lin, Lingxiao Tang, Yuyu Luo, Bang Liu, and Chenglin Wu. 2026. ReCode: Unify Plan and Ac- tion for Universal Granularity Control. Preprint, arXiv:2510.23564. Ran Zhang and Steffen Eger. 2024. LLM-based multi- agent poetry generation in non-cooperative environ- ments. Preprint, arXiv:2409.03659. Weizhi Zhang, Yangning Li, Yuanchen Bei, Junyu Luo, Guancheng Wan, Liangwei Yang, Chenx- uan Xie, Yuyao Yang, Wei-Chieh Huang, Chunyu Miao, Henry Peng Zou, Xiao Luo, Yusheng Zhao, Yankai Chen, Chunkit Chan, Peilin Zhou, Xinyang Zhang, Chenwei Zhang, Jingbo Shang, and 4 others. 2025. From Web Search towards Agentic Deep Re- search: Incentivizing Search with Reasoning Agents. Preprint, arXiv:2506.18959. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, and 1 others. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information pro- cessing systems, 36:46595â46623. Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. 2025. DeepResearcher: Scaling Deep Research via Re- inforcement Learning in Real-world Environments. Preprint, arXiv:2504.03160. A Theoretical Analysis of Convergence In this section, we provide a formal analysis of the convergence properties of CogGenâs parallel- recursive architecture. We model the report gener- ation process as a discrete dynamical system and analyze how the proposed Reviewer Gating Mech- anism acts as a monotonic filter, promoting con- vergence toward a stable local optimum. Under the premise of noisy LLM judgments, this mecha- nism is best understood as an empirically effective heuristic rather than a strict theoretical guarantee. A.1 System Modeling LetSbe the state space of all possible report drafts. A stateS t â Sat iterationtis defined by the tuple(O (t) ,C (t) ), representing the current outline and content. We define an Inconsistency Energy Function E :S âR âĽ0 , which quantifies the total logical conflict and quality deficit within a report. E(S t ) = N X i=1 Loss local (c i ) + Îť X i,j Conflict(c i ,c j ) (6) whereLoss local quantifies the quality deficit of a single section, andConflictrepresents logical con- tradictions between sectionsiandj. A perfect report corresponds to a stateS â whereE(S â )â 0. A.2 Convergence of Deferred Resolution The core challenge in recursive writing is Con- textual Oscillation, where a local repair in sec- tioniincreases the conflict with sectionj, causing E(S t+1 ) > E(S t )and leading to limit cycles (in- finite loops). CogGen addresses this via the De- ferred Resolution Strategy and Global Review Gating. Proposition 1 (Convergence under Idealized Gating). The CogGen generation process con- verges to a local optimum if the Reviewer AgentA r enforces a strict energy descent condition. Proof Sketch. In the parallel phase, the Writer generates a candidate set of updatesâS. The Re- viewerA r does not accept these updates individu- ally. Instead, it evaluates the aggregated next state S Ⲡt+1 . The Gating Mechanism (Eq. 7) accepts the transition S t â S t+1 if and only if: Q(S Ⲡt+1 )â Q(S t )⼠ξ(7) whereQis the quality score estimated by the LLM (an inverse proxy for EnergyE) andÎľ > 0is a min- imum improvement threshold. Since the state space of meaningful reports is finite and bounded, and the quality scoreQis bounded from above (e.g., by the maximum context window capacity or log- ical completeness), a strictly increasing sequence Q(S 0 ),Q(S 1 ),...must converge to a fixed point where no further improvement⼠ξis possible. At this point, the system terminates. A.3 Complexity Advantage Unlike serial backtracking, which suffers from worst-case exponential complexity due to cas- cading edits (O(k N )in naive recursive repair), CogGenâs parallel update dampens the complex- ity. By calculating updates for all defect nodes simultaneously, CogGen approximates the gradient descent direction of the Energy functionEover the entire report structure. Assuming the decoupling of sections allows for independent convergence rates, the time complexity is dominated by the slowest converging section rather than the sum of all revi- sions: T CogGen â max i (m i )¡ T step (8) wherem i is the number of revisions for sectioni. This represents a significant speedup over the serial cumulative time P m i ¡ T step . Empirical Validation. These theoretical conver- gence properties are corroborated by the execution statistics presented in Appendix B. Specifically, the low Global Restructure Rate (16.0%) and the rapid generation latency (3.61 min) detailed in Table 7 validate that the parallel architecture effectively suppresses worst-case oscillation, aligning with our complexity analysis. B Experimental Analysis In this section, we analyze the computational effi- ciency of CogGen. We first provide a formal speci- fication of the parallel execution mechanism, then benchmark the generation latency against baseline models (Table 6), and finally provide a granular decomposition of CogGenâs internal execution to explain the source of latency and validate the sys- temâs architectural stability (Table 7). B.1 Formal Specification of Parallel Execution This subsection provides the formal specification of CogGenâs parallel micro-cycle execution, includ- ing the write isolation constraints and knowledge base synchronization protocol referenced in Sec- tion 3.3. Write Isolation Constraint.Each parallel thread Thread s operates as a read-only observer of the global outlineO (t) and all other sectionsâ content C (t) j̸=s . No thread may modify the outline or any other sectionâs content during execution. This in- variant is enforced architecturally: threads receive a frozen copy ofO (t) at the start of each macro- iteration, eliminating race conditions by construc- tion. Hierarchical Knowledge Base Protocol. The knowledge baseKis partitioned into a Global TierK g and a Local TierK s . The global tier is a shared, immutable snapshot constructed during macro planning; all threads read from the same K g . The local tier is a thread-local cache where Thread s stores evidence retrieved during its micro- cycle retrieval phase, invisible to other threads. The effective knowledge available toThread s is there- foreK eff (s) = K g ⪠K s , whereK s ⊠K s Ⲡ= â fors̸= s Ⲡ. This isolation prevents irrelevant noise from propagating between unrelated chapters. Execution Sequence. The parallel micro-cycle proceeds through three phases. First, in the Dis- patch and Parallel Planning phase, the macro con- troller broadcastsO (t) andK g to all threads. Each Thread s independently performs targeted retrieval and generates a section-level plan foro s â O (t) , populating its local cacheK s . Second, a syn- chronous Coarse-Grained Plan Aggregation step consolidates all section-level plans, performing cross-section deduplication and boundary adjust- ment to eliminate redundancy before writing be- gins. This lightweight, structure-level consistency pass ensures that parallel plans do not overlap or conflict at the outline level. Third, in the Parallel Writing phase, eachThread s composes the content c s based on its consolidated plan, executing the recursive WriteâReview micro-loop. A barrier syn- chronization ensures all threads complete before the unified draftC (t) =c s |âs is assembled. Two-Tier Consistency Architecture. Once the complete draft is available, the ReviewerA r per- forms a Fine-Grained Global Reviewâa holistic, content-level evaluation that detects cross-section logical conflicts, factual inconsistencies, and struc- tural imbalances that the coarse-grained plan aggre- gation cannot captureâand produces the feedback signalâ (t) . The transitionO (t) â O (t+1) is ac- cepted only if the quality improvement exceeds the thresholdÎľ(Eq. 7). This two-tier designâcoarse- ModelTime (min) Linear Models STORM1.54 CO-STORM3.55 Recursive Models Multimodal DeepResearcher10.46 WriteHere14.75 CogGen (Ours)20.50 Table 6: Efficiency comparison on the OWID dataset (N = 40). MetricValue Time & Latency Retrieval Duration16.89 min (82.4%) Generation Duration3.61 min (17.6%) Retrieval Latency / req78.05 s Generation Latency / req5.48 s Resource Allocation Avg. Costâ $4.80 Total Tokens5.01 M - Retrieval Phaseâ 80% - Generation Phaseâ 20% Execution Dynamics Plan Modifications2.39 Content Modifications0.43 Zero-Shot Success71.1% Restructure Rate16.0% Table 7: Internal execution statistics of CogGen. Data represents averages from the OWID dataset (N = 40). grained aggregation before writing and fine-grained review after writingâensures that no partial state is ever observed by the Reviewer, enabling deter- ministic conflict resolution while minimizing re- dundant generation effort. B.2 Latency Analysis Table 6 compares the average generation time across five report generation frameworks. We ob- serve a distinct stratification in temporal perfor- mance, which correlates with the depth of infor- mation processing and the retrieval strategies em- ployed. Retrieval Pipelines and Fidelity. While all frameworks in our evaluation utilize the Tavily Search as the unified retrieval source, their post- retrieval processing strategies diverge significantly to align with their respective architectural goals. Snippet-based Processing. Baselines such as STORM and WriteHere are designed to optimize for response speed. They typically ingest search snippets or RAG-retrieved chunks directly. While efficient, we argue that for long-form report gen- eration, relying solely on snippets carries the risk of contextual fragmentation, where disconnected text segments may induce logical inconsistencies or hallucinations during synthesis. Full-Content Summarization.In con- trast, CogGen explicitly implements a Crawler- Summarizer Pipeline (reading full web pages and summarizing via LLM), aligning with the techni- cal framework of deep research agents like Tongyi DeepResearch (Li et al., 2025a). We treat this com- putationally intensive step as a necessary âDenois- ing and Verificationâ layer. By digesting the com- plete document context before synthesis, the model filters out irrelevant noise and ensures better logical coherence, effectively mitigating the hallucination risks inherent in snippet-stitching approaches. Impact on Quality Assessment. Crucially, this comprehensive ingestion strategy does not artifi- cially inflate the structural or multimodal evalua- tion metrics (e.g., Organization, Alignment) used in CLEF. Instead, its primary function is to miti- gate hallucinations. By ensuring that the model reasons over verified summaries rather than frag- mented snippets, we guarantee that the high scores achieved in the âDepthâ dimension (Table 3) reflect genuine analytical capability rather than plausible- sounding fabrications. This ensures a rigorous and valid quality comparison where CogGenâs advan- tage stems from its recursive architecture, not just data quantity. Latency Attribution and Architectural Speed. Table 6 indicates that CogGenâs total latency (20.50 min) is higher than the snippet-based baselines. It is crucial to note that 82.4% of this time is allocated to the heavy Ingestion Phase (full-page reading and summarization), a deliberate design choice to prioritize information fidelity over raw speed. Most importantly, when isolating the Reason- ing and Generation Phase (Table 7), CogGen com- pletes the complex multimodal planning and writ- ing in only 3.61 minutes. This confirms that our Gated Parallelism mechanism effectively solves the bottleneck of recursive generation, achieving a throughput significantly higher than serial recur- sive baselines like WriteHere (14.75 min). Future deployments could mitigate retrieval latency by employing specialized lightweight summarization models instead of general-purpose LLMs. B.3 Internal Dynamics and Stability To validate the Deferred Update mechanism pro- posed in Section 3.3, we analyze the internal be- havioral statistics of CogGen. Table 7 details the resource consumption and modification patterns. Planning Flux vs. Writing Stability.The statis- tics reveal a functional decoupling between plan- ning and writing. The Planner exhibits high activity (2.39 revisions/section), absorbing the uncertainty of the task. In contrast, the Writer demonstrates high stability (0.43 revisions/section) with a 71.1% zero-shot success rate. The 5.6:1 ratio between plan and content revisions provides empirical evi- dence that the hierarchical architecture effectively transforms a complex reasoning problem into a deterministic execution task. Crucially, this stability does not imply rigidity. The Global Restructure rate (16.0%) indicates that while the local writing link prioritizes efficiency optimization, the global planning link maintains the flexibility to adapt to logical conflicts discov- ered during execution. This hierarchical dynamism ensures that the system avoids the âtunnel visionâ typical of linear models while minimizing the la- tency cost of full recursion. B.4 Backward Restructuring Analysis To provide concrete evidence of the backward re- structuring mechanism described in Section 3.2, we analyze restructuring events observed across the evaluated reports. Frequency and Outcomes.Across all evaluated reports, 13.3% of outline modifications involve backward restructuringâcases where downstream content discoveries trigger retroactive changes to the global outline. We manually examined all ob- served backward restructuring events and found no harmful updates. All cases involved structural optimizations such as eliminating cross-section re- dundancy and adjusting section boundaries, with consistent Reviewer decision direction. Representative Example. In a report on âWhat are the safest and cleanest sources of energy?â, the Reviewer identified content overlap between §2.1âs comprehensive ranking and Chapter 6âs summary synthesis during the macro-cycle review, triggering backward restructuring. Table 9 presents the origi- nal outline and the Plannerâs targeted modification instructions. ModelOrganizationDepthRelevanceAlignmentSynergyAvg. Score Evaluator I: Doubao-Seed-1.6 (Judge) Gemini Deep Research (Ref)0.50000.50000.50000.51500.50000.5030 Multimodal DeepResearcher0.39380.35360.36410.23000.24330.3170 WriteHere 0.54660.53820.52610.53520.50050.5293 CogGen (Ours)0.55910.55280.55000.65480.67620.5986 Evaluator I: Claude-Sonnet-4 (Judge) Gemini Deep Research (Ref)0.50280.50280.50000.58330.64940.5477 Multimodal DeepResearcher0.30800.29910.30800.20170.19670.2627 WriteHere 0.56100.56090.54730.52980.55620.5510 CogGen (Ours)0.54740.57150.53340.71810.65720.6055 Table 8: Robustness Analysis on WildSeek Dataset across Evaluators. Comparison of model performance when evaluated by different judge models: Doubao-Seed-1.6 (top) and Claude-Sonnet-4 (bottom). Bold highlights the best result, and underlinedmarks the second best. StageContent Before 2.1 Comparative Ranking of Lifecycle GHG Emissionsâ Analyze lifecycle GHG emissions for major energy sources using quantified values for coal, oil, gas, nuclear, wind, solar. . . After⢠Trim paragraphs that broadly summarize which sources are âdirtiestâ or âcleanestââleave detailed ranking and synthesis for §6. ⢠Keep detailed lifecycle GHG data, methodology, and regional/technological variability analysis. ⢠Remove summary statements duplicating §6âs synthesis. Table 9: Backward restructuring example: the Plannerâs revision of §2.1 to eliminate cross-section redundancy with Chapter 6. After modification, Section 2.1 retained detailed lifecycle emission data and methodological anal- ysis, while comprehensive conclusions were de- ferred to the final chapter, eliminating cross-section redundancy. C Detailed Evaluation C.1 Evaluation Across Different Models In the main text, we adopt GPT-5 as the primary evaluation judge owing to its superior reasoning capabilities and strong alignment with human pref- erences. To mitigate potential biases induced by the choice of a single evaluation model and to ver- ify the cross-model robustness of our results, we further conducted experiments on the WildSeek dataset using two distinct state-of-the-art LLMs as alternative judges: Doubao-Seed-1.6 and Claude- Sonnet-4. The comparative evaluation results un- der the CLEF framework are presented in Table 8. It is important to note that the model employed in our generation process (CogGen) is completely in- dependent of these judge models, ensuring a blind evaluation setting. As shown in Table 8, while the absolute scoring distributions vary between judges (e.g., Claude- Sonnet-4 tends to assign higher baseline scores to the reference), the relative performance trends remain highly consistent. CogGen maintains the highest Overall Average Score across all evalua- tors. Notably, in critical multimodal metrics such as Alignment and Synergy, CogGen consistently outperforms baselines by a significant margin re- gardless of the evaluator used. These results con- firm that our methodâs superiority is intrinsic to the generated content quality and is robust to the variations in evaluation models. C.2 Human Comparative Evaluation To rigorously validate CogGenâs effectiveness, we conducted a blinded head-to-head human evalua- tion on WildSeek, comparing against two baselines: (1) Multimodal DeepResearcher (MMDR), a mul- timodal baseline using a linear workflow; and (2) Gemini Deep Research, a proprietary commercial system, to benchmark overall performance. C.2.1 Setup Evaluation Protocol. We evaluated all 20 Wild- Seek queries without sampling to eliminate selec- tion bias. A blinded annotator assessed each re- port pair across four dimensions: Overall Quality, WriteHere Multimodal DeepResearcher CogGen The Role of Seroprevalence Studies in Calibration and Validation Seroprevalence studies, which measure the proportion of individuals with antibodies to SARS- CoV-2, provide a direct estimate of cumulative infection rates. These surveys are invaluable for calibrating and validating epidemiological models. By comparing model-based estimates of infections to seroprevalence data, researchers can assess the accuracy of their assumptions about under-ascertainment and adjust parameters as needed. For example, in Spainâs ENE-COVID study, nationwide seroprevalence was found to be several times higher than the number of confirmed cases, prompting modelers to revise their estimates of detection rates and infection- to-case ratios . Similarly, seroprevalence data from New York City revealed that actual infections were far greater than official counts, validating model-based projections and informing public health responses . These approaches are invaluable for calibrating models and highlighting regional variation in ascertainment gaps. For example, by April 2021, the global seroprevalence attributable to infection was estimated at 21%âalmost exactly matching IHMEâs modeled cumulative incidence of 24%. However, crucial regional differences remain; some regions (AFR, WPR) lacked reliable seroprevalence-to-case ratios until recent studies began to fill these gaps. Visualization 3 For instance, Belgiumâs post-peak decline in seroprevalence and Japanâs increase during later waves underscore how sero-surveys reveal the timing and completeness of epidemic surveillance. Figure3.1: Confirmed COVID-19 Cases vs. Sero-Prevalence- Based Estimated Infections Over Time(7[17]) As Figure 3.1 demonstrates, confirmed cases consistently lag behind sero-prevalence-based estimates, with the gap most pronounced in regions and phases of limited testing. The pattern reinforces the necessity of supplementing case data with empirical infection surveys for accurate epidemic reconstruction. Figure 3: Qualitative Comparison of Cross-Modal Alignment Performance: The left panel displays the output of the baseline model WriteHere; the middle panel presents the generated results of Multimodal DeepResearcher; and the right panel shows the output of our proposed CogGen method. We adopt a color-coded highlighting approach to mark the correspondences between textual content and visual elements. DimensionW/T/LWin% Overall Quality18/1/190.0 â Content Depth 19/0/195.0 â Visual-Text Alignment 16/2/280.0 â Multimodal Synergy16/2/280.0 â Table 10: Human evaluation: CogGen vs. MMDR (N =20). W/T/L: Win/Tie/Loss. â : p<0.01. DimensionW/T/LWin% Overall Quality15/1/475.0 â Visual-Text Alignment 16/0/480.0 â Multimodal Synergy16/1/380.0 â Content Depth 10/3/750.0 Table 11: Human evaluation: CogGen vs. Gemini (N =20). â : p<0.05; â : p<0.01. Alignment, Synergy, and Depth. Statistical signifi- cance was assessed using the Wilcoxon signed-rank test (ties excluded). C.2.2 Results Tables 10 and 11 present the comparative results. AVR Mechanism Validation.Table 10 demonstrates CogGenâs substantial advantage over MMDR across all dimensions (win ratesâĽ80%). The 95% win rate in Depth validates our multi- modal reasoning framework, while consistent 80% wins in alignment and synergy dimensions em- pirically confirm AVRâs effectiveness in bridging the reasoning-rendering semantic gap compared to MMDRâs implementation. Gemini Comparison.Compared with the Gemini Deep Research (Gemini) (see Table 11), CogGen achieves a statistically significant advan- tage in both Overall Quality (75% win rate,p < 0.05) and Multimodal Dimension (80% win rate,p < 0.01). We draw two core findings: (1)Multimodal Advantage: CogGenâs AVR mech- anism enables precise, context-aware chart place- ment; while Gemini generates abundant tables, they often lack contextual relevance. (2)Reasoning Par- ity: CogGen ties with Gemini (50% win rate) in the Content Depth dimension. This demonstrates that the hierarchical recursive framework proposed in our study not only excels in multimodal fusion performance, but also matches the reasoning capa- bilities of proprietary commercial systems. C.3 Case Study Due to space constraints in the main text, we place the qualitative case comparison in the appendix, as illustrated in Figure 3. We compared three frame- worksâWriteHere, Multimodal DeepResearcher, and CogGenâregarding their descriptive perfor- mance on seroprevalence-based approaches. Em- pirical examples demonstrate that WriteHere gener- ates text-only content, with no quantitative results included in its case descriptions. Multimodal Deep- Dimension & CTML PrinciplesEvaluation Focus & Cognitive Goal Control Dimensions: General Quality D1: Information OrganizationHierarchical structure. Evaluates if headings and layout guide attention effec- tively via signaling. (Signaling, Segmenting)(Extraneous Loadâ) D2: Content Depth and Insight Causal explanations. Assesses whether content promotes deep reasoning and schema construction over fact stacking. (Concreteness, Schema Construction)(Germane Loadâ) D3: Content Relevance and AdaptationAppropriate complexity. Ensures content excludes distracting details and matches user intent. (Coherence, Pre-training, Personalization)(Extraneousâ, Intrinsic Managed) Core Dimensions: Multimodal Integration D4: Visual-Text AlignmentTight spatial/semantic integration. Assesses whether elements are physically and semantically close to reduce split-attention effects. (Spatial Contiguity)(Extraneous Loadâ) D5: Multimodal SynergyComplementarity beyond text. Checks if visuals provide unique information gain (e.g., trends) rather than decoration. (Multimedia, Redundancy, Image)(Extraneousâ, Germaneâ) Table 12: Detailed definitions of CLEF dimensions, mapped to CTML Principles and cognitive load targets. Researcher produces content integrating text and graphics; however, the textual component lacks an- alytical depth, and there is no logical correlation between the images and text, which instead dis- rupts the normal reading flow. In contrast, CogGen, the method proposed in this study, conducts a cross- sectional data comparison between Japan and Bel- gium, employs line charts to intuitively visualize the developmental trends, and achieves tight inte- gration of text and graphics along with targeted in-depth analysis. C.4 Bootstrap Significance Analysis To rigorously assess statistical significance, we conducted Bootstrap analysis (B=10,000) on the CLEF evaluation results. CogGen is the only sys- tem whose overall score shows no significant dif- ference from the human reference level (p=0.88, 95% CI fully covering 0.5), whereas all baselines fall significantly below (p<0.001). The advantage is most pronounced on the multimodal dimensions (Alignment and Synergy), where CogGen outper- forms the strongest baseline WriteHere by over 0.09 points (p<0.001). C.5 Cross-Domain Evaluation To verify that CogGenâs advantages are not overfit to the original OWID topic distribution (concen- trated in Health & Medicine at 32.5% and Eco- nomics & Development at 17.5%), we collected 10 additional multimodal reports spanning previ- ously underrepresented domains including Democ- racy/Governance, Social Media/Digital Technol- ogy, Immigration/Demographics, Financial Tech- nology, Media/Public Perception, and Gender/De- mography. Table 13 presents the evaluation results. The results are consistent with the main exper- iment trends: CogGen maintains the overall lead (Avg. 0.486), with particularly significant advan- tages on multimodal dimensions (Alignment and Synergy). This confirms that the hierarchical re- cursive architecture generalizes across diverse do- mains. D CLEF: Cognitive Load Evaluation Framework Details D.1 Theoretical Foundation CLEF is grounded in two complementary theories: Cognitive Load Theory (CLT) CLT identifies three types of cognitive load: intrinsic load (con- tent difficulty), extraneous load (presentation bur- den, to be minimized), and germane load (schema construction effort, to be maximized) (Sweller, 1994). Cognitive Theory of Multimedia Learning (CTML) Mayerâs CTML operationalizes cogni- tive principles into measurable design dimensions. CLEF maps these principles to evaluation metrics to assess cognitive burden reduction (Mayer, 2005). D.2 Evaluation Dimensions We map the evaluation dimensions to specific CTML principles and CLT goals. Table 12 details the evaluation focus for each dimension. ModelOrganizationDepthRelevanceAlignmentSynergyAvg. Score CogGen (Ours)0.5000.5900.4720.4490.4210.486 Multimodal DeepResearcher0.4030.4750.3750.1790.1700.320 WriteHere0.4890.5470.5110.4380.3780.473 Table 13: Content quality evaluation on 10 additional cross-domain reports. Scores are CLEF Relative Advantage. D.3 Complete Mapping to CTML Principles Table 12 presents the primary CTML principles associated with each evaluation dimension. To pro- vide a comprehensive view, Table 14 presents the complete mapping from all 14 CTML principles to CLEF dimensions, clarifying coverage and scope. Coverage AnalysisCLEFâs five dimensions sys- tematically operationalize 11 of the 14 CTML principles. Three principles (Modality, Tempo- ral Contiguity, Voice) are excluded as they specif- ically address dynamic multimedia (audio/video synchronization) and are not applicable to static text-visual reports. The framework comprehen- sively addresses all three CLT load types: mini- mizing extraneous load through D4, D1, and D2; managing intrinsic load via D3; and promoting germane load through D5 and D2. D.4 Scoring Mechanism Pairwise Comparative Evaluation Following best practices (Du et al., 2025), GPT-5 simultane- ously evaluates both the model report and a refer- ence report. Relative Advantage Score For each dimension i, the relative advantage score is calculated as: R i = S (i) model S (i) model + S (i) ref â [0, 1](9) whereR i > 0.5indicates the model report outper- forms the reference. The final score is the average across all dimensions: R final = 1 5 5 X i=1 R i (10) D.5 Implementation Prompt StructurePrompts are structured to mit- igate âLost in the Middleâ effects (Liu et al., 2024): (1) evaluation rubric (as defined in Table 12); (2) interleaved text-image content of both reports; (3) holistic comparative instructions. Images are en- coded in base64 to leverage GPT-5âs native multi- modal capabilities. E Factuality Evaluation To quantify CogGenâs factual reliability, we con- ducted both automated and human-verified evalua- tions on the WildSeek dataset. E.1 Evaluation Methodology Automated Citation Evaluation. We collected all citations from each systemâs reports across 20 WildSeek queries (11,291 total citations). For each citation, we crawled the source URL and used an LLM to judge the relevance of the cited content to the corresponding statement, computing Citation Precision. Human Claim-Level Verification.We sampled 5 reports from each system and decomposed the most claim-dense paragraphs into 148 atomic claims. Human annotators independently verified each claim via web search, measuring two met- rics: Supported Rate (proportion of claims with supporting web evidence) and Citation Accuracy (proportion of cited sources that actually contain the claimed content). E.2 Results CogGen achieves the highest scores across all three factuality metrics: Citation Precision of 0.72 (vs. WriteHere 0.69, Gemini 0.60), human-verified Sup- ported Rate of 76.3% (vs. 72.7%, 60.5%), and Cita- tion Accuracy of 55.3% (vs. 54.5%, 44.2%). These results demonstrate competitive factual reliability even without dedicated optimization for this dimen- sion. E.3 Ingestion Strategy Ablation To disentangle the contributions of retrieval strat- egy (ingestion) and recursive architecture, we re- placed CogGenâs full-text summarization strategy with lightweight snippet retrieval. The two configu- rations differ only in the retrieval stage; the writing model receives context in an identical format. Switching from full-text summarization to snip- pet retrieval yields nearly identical CLEF scores CTML PrincipleCLEF DimensionMapping Rationale Principles Directly Evaluated by CLEF 1. Multimedia PrincipleD5Assesses whether text-visual combinations provide synergistic informa- tion gain beyond text alone. 2. Modality PrincipleN/AConcerns audio vs. text; not applicable to static multimodal reports. 3. Redundancy PrincipleD5Evaluates whether visuals complement text rather than merely repeating it verbatim. 4. Spatial ContiguityD4Measures spatial proximity between related text and visual elements to reduce split-attention. 5. Temporal ContiguityN/AConcerns synchronization in dynamic media; not applicable to static reports. 6. Coherence PrincipleD3 Checks whether content excludes extraneous, distracting, or irrelevant information. 7. Interactivity PrincipleN/A Concerns learner-controlled pacing; not applicable to static report evalua- tion. 8. Signaling PrincipleD1Evaluates use of headings, highlighting, and structural cues to guide attention. 9. Segmenting PrincipleD1 Assessed through hierarchical organization and logical content chunking. 10. Pre-training PrincipleD3Indirectly evaluated via content adaptation to user expertise level. 11. Personalization PrincipleD3Considered in evaluating whether content tone and complexity match user intent. 12. Concreteness PrincipleD2Assesses use of examples, analogies, and concrete instantiations in expla- nations. 13. Voice PrincipleN/AConcerns audio narration quality; not applicable to text-based reports. 14. Image PrincipleD5Evaluates whether images serve functional (not decorative) purposes. Cognitive Load Theory (CLT) Integration Intrinsic LoadD3 Managed through appropriate content complexity matching user expertise. Extraneous LoadD4, D1, D3Minimized via spatial integration (D4), clear structure (D1), and coher- ence (D3). Germane LoadD5, D2Enhanced via meaningful visual integration (D5) and deep explanations (D2). Table 14: Complete mapping from Mayerâs 14 CTML principles and 3 CLT load types to CLEFâs 5 evaluation dimensions. Principles marked N/A are not applicable to static multimodal report evaluation. (0.4992 vs. 0.5019) but sharply reduces the Sup- ported Rate from 76.3% to 50.0%, while generation time drops from 20.50 to 6.62 minutes. This re- veals a clear separation of concerns: CLEF scores are nearly identical, indicating that CogGenâs con- tent quality advantage stems from the hierarchical recursive architecture and AVR mechanism, not the retrieval strategy. However, the Supported Rate drops sharply, confirming that the full-text sum- marization pipeline is critical for factual accuracy. With 82.4% of total latency attributable to the re- trieval stageârecursive reasoning itself requires onlyâź3.6minutesâusers can flexibly choose be- tween a factuality-first mode (20 min) and a speed- first mode (7 min) depending on the use case. F Visualization Implementation Details This appendix provides a comprehensive analysis of the visualization generation module in CogGen, detailing the Abstract Visual Representation (AVR) design, the rendering pipeline, architectural trade- offs compared to related work, and statistical vali- dation on the OWID dataset. F.1AVR-based Decoupled Rendering Pipeline As introduced in Table 1 of the main text, the Ab- stract Visual Representation (AVR) serves as the intermediate bridge between narrative intent and visual execution. The generation process follows a strict pipeline: the Planner determines the chart intent, the Writer generates the AVR structure, and the Render Agent translates AVR into executable code. AVR Field Structure. To ensure generative sta- bility, the AVR schema is divided into mandatory and optional fields: â˘Fixed Fields (Mandatory): Required for every visualization to define the core in- tent.These includeTitle,Chart_Type, Data_Source, and Purpose. â˘Dynamic Fields (Optional):Context- dependent fields such asX_AxisandY_Axis definitions, which are only generated when the specifiedChart_Typerequires coordinate mapping (e.g., Bar Charts) and are omitted for types like Pie Charts or Flowcharts. Rendering Technology Stack. While LLMs in- creasingly demonstrate the ability to generate raw HTML/CSS directly, we deliberately constrain the Render Agent to target specific high-level visualiza- tion libraries: Mermaid.js and Apache ECharts. â˘Implementation Strategy: Rather than per- mitting the Render Agent to freely halluci- nate HTML structuresâwhich often leads to inconsistent styling and broken layoutsâthe agent generates configuration code for these libraries. ⢠Execution Environment: The rendering oc- curs in a browser-based environment. Lever- aging established frontend libraries ensures interactive, aesthetically consistent, and func- tionally robust charts while significantly low- ering the coding capability requirement for the LLM. F.2 Cognitive Load Trade-off and Comparison Our design philosophy centers on minimizing the Dual-Task Interference for the Writer agent. We explicitly trade off granular control for semantic simplicity. Comparison with Multimodal DeepResearcher. Existing systems like Multimodal DeepResearcher (MMDR) adopt a âTwo-Stageâ rendering strategy using a placeholder known as FDV (Formal De- scription of Visualization). The FDV is designed to describe every visual detail, including style, color, and layout, with high precision. â˘The MMDR Limitation: Our empirical ob- servations indicate that such verbose place- holders impose a substantial cognitive load on the Writer agent. Attempting to perfect visual specifications distracts the model from its pri- mary task of narrative construction, leading to degradation in text quality. ⢠The CogGen Advantage: By offloading styling decisions to the standard themes of ECharts and Mermaid, the AVR allows the MetricValue Generation Performance Total Reports40 Requested Visualizations258 Successfully Generated248 Success Rate96.12% Avg. Visualizations per Report6.45 Type Diversity Distinct Chart Types22 Primary Categories4 Table 15: Visualization generation statistics on the OWID dataset. Writer to focus solely on data and intent. This âlightweightâ representation reduces cognitive overhead, preventing the quality dip observed in MMDR. Quantitative Comparison.To quantify the cog- nitive cost difference, we measured the average token count per visualization placeholder across 50 reports. AVR averagesâź133tokens per figure (measured over 339 blocks), while FDV averages âź773tokens (measured over 252 blocks)âa 5.8Ă difference. This reduction directly reflects the sep- aration of concerns: AVR answers âwhat to show and whyâ while delegating âhow to drawâ to the dedicated Render Agent. Post-Rendering Data Verification Pipeline.As discussed in Section 5.4, AVRâs decoupled nature enables a Post-Rendering Audit, which is architec- turally difficult in FDVâs monolithic pipeline. In CogGen, this module operates by parsing the in- termediate ECharts JSON generated by the Render Agent and cross-checking the exact coordinate data points against the original source values retrieved in the Knowledge BaseK. This verification-in-the- loop mechanism is responsible for the significant drop in hallucination rates detailed in Table 5 of the main text. F.3 Statistical Analysis of Generated Visualizations To validate the effectiveness of our multimodal report generation system, we conducted a compre- hensive statistical analysis on the visualization out- puts from the OWID dataset (N = 40). Table 15 summarizes the key quantitative metrics. High Generation Reliability. The system achieved a 96.12% success rate across 258 visual- Chart TypeCount%Cumulative Statistical Charts (46.9%) Bar Chart6926.7426.74 Line Chart3915.1241.86 Area Chart135.0446.90 Relational & Process (25.6%) Flowchart3814.7361.63 Heatmap72.7164.34 Pie Chart62.3366.67 Timeline62.3368.99 Scatter Plot51.9470.93 Sankey41.5572.48 Geographic & Structural (20.2%) Map186.9879.46 Diagram176.5986.05 Infographic103.8889.92 Matrix83.1093.02 Specialized (7.0%) Table62.3395.35 Roadmap41.5596.90 Others (6 types)83.10100.00 Table 16: Distribution of generated chart types across functional categories. ization requests, demonstrating robust cross-modal generation capability. Each report contains an average of 6.45 visualizations, indicating that the system effectively integrates visual elements to support textual content. This high reliability validates the architectural design of our multimodal generation pipeline. Chart Type Distribution.Table 16 presents the distribution of generated chart types across func- tional categories. The system demonstrates strong diversity, producing 22 distinct chart types span- ning statistical analysis, process visualization, ge- ographic mapping, and specialized structural dia- grams. Dominance of Statistical Charts. As shown in Table 16, basic statistical charts (bar, line, area) account for 46.9% of all visualizations, consistent with the data-driven nature of analytical reports. The high prevalence of bar charts (26.74%) reflects their versatility in comparative analysis, while the frequent use of line charts (15.12%) indicates a focus on trend visualization. ProminenceofProcessVisualization. Flowcharts rank third at 14.73%, a notably high proportion for non-statistical charts. This suggests that the generated reports emphasize logical relationships and procedural explanations alongside raw data presentation. The combined relational and process chart category (25.6%) demonstrates the systemâs capability to handle complex structural reasoning beyond simple data plotting. Multimodal Type Diversity. Beyond basic sta- tistical charts, the system generates a rich variety of specialized visualizations including geographic maps (6.98%), structural diagrams (6.59%), in- fographics (3.88%), and matrices (3.10%). This demonstrates the systemâs ability to select appro- priate visual encodings for diverse analytical con- textsâfrom spatial data (maps) to conceptual rela- tionships (diagrams) to decision frameworks (matri- ces). The presence of 22 distinct chart types across 4 functional categories validates the systemâs mul- timodal reasoning capability. Rendering Technology Distribution. The sys- tem employs a dual-technology stack: ECharts handles 81.9% of visualizations (primarily data- driven charts and maps), while Mermaid man- ages 18.1% (flowcharts and architectural dia- grams). This division aligns well with each li- braryâs strengthsâECharts for quantitative visu- alization and Mermaid for declarative diagram syn- taxâresulting in efficient and appropriate technol- ogy allocation. Coverage and Concentration. The type distri- bution exhibits a natural concentration pattern: the top 10 chart types cover 84.9% of all visualizations, indicating a stable set of core visualization patterns. Simultaneously, the presence of specialized types (accounting for 15.1% of charts) demonstrates the systemâs flexibility to adapt to domain-specific an- alytical needs. This balance between standardiza- tion and specialization reflects effective alignment between the systemâs multimodal generation capa- bility and the diverse requirements of analytical report writing. G OWID Dataset Construction We constructed our evaluation dataset from Our World in Data (OWID), 1 a widely-cited platform for data-driven research reports. The construction involved three stages: web scraping, quality filter- ing, and format standardization. 1 https://ourworldindata.org G.1 Data Collection We developed an automated web scraper to col- lect reports from OWIDâs publication archive (De- cember 2016âSeptember 2025). The scraper ex- tracts complete report content (title, publication date, authors, main text, embedded visualizations) and implements politeness controls (1â2 second request delays, automatic retry mechanisms). This process collected 399 reports across diverse top- ics including health, environment, economics, and social issues. G.2 Filtering To focus on substantive research reports and ex- clude announcements or atypical content, we applied the following criteria: Content length: 15,000â60,000 characters; Word count:âĽ2,500 words; Visualizations: 3â15 images per report; Excluded keywords like âAnnouncingâ, âWelcom- ingâ. The minimum requirements ensure sufficient content for meaningful evaluation, while maximum thresholds remove edge cases (e.g., comprehensive handbooks, image repositories). The visualization constraint focuses on typical research reports with substantive multimodal integration. Furthermore, we verified that the retained reports are free of sensitive personally identifiable information (PII). After filtering, 40 high-quality reports remained (10.04% retention rate). G.3 Format Standardization Reports were standardized for evaluation use. HTML content was converted to Markdown for- mat preserving document structure (headings, para- graphs, lists). Crucially, visualization references in text were mapped to their corresponding image files, maintaining the spatial and semantic relation- ships between text and visuals. This image-text alignment is essential for evaluating multimodal integration quality. Metadata (source, publication date, content statistics) was preserved for repro- ducibility. G.4 Dataset Statistics The compiled dataset comprises 40 reports, averag- ing 3,625 words and 7.9 visualizations per report. Reports span diverse topics with substantial multi- modal content, providing a challenging testbed for automated report generation systems. H Prompt In this section, we provide the evaluation prompts for our framework, including a template and met- rics across five dimensions. These prompts were also used by human evaluators. Due to the large number of prompts required for individual agents and intermediate processes in CogGen, the system prompts will be released along with the code. Prompt for Evaluation Template query_sectionrubric ---Separator: Below are two reports to be compared on the same dimension (including text and charts)--- report1 report2 ---Separator: End of two report contents--- [Evaluation Task] You need to **simultaneously** evaluate Report A (Model Report) and Report B (Reference Report) on the "dimension_name" dimension, and provide relative advantage judgment. **Evaluation Method**: 1. Read both reports completely to form an overall quality impression 2. Please understand the intent of the user question and the purpose of the report, and consider whether the reportâs organization matches these intents and purposes 3. Determine which score range description (1-5 points) each reportâs overall performance is closer to 4. Score based on overall quality level Please refer to the description of each score level (1-5 points) in the [Scoring Rubric] section of the rubric above, and determine: - Which score range (integer between 1-5) Report Aâs overall performance on this dimension is closer to - Which score range Report Bâs overall performance on this dimension is closer to - Which one is overall better on this dimension, and what are the reasons [Output Requirements] Please output the comparison results in JSON format (do not include markdown code block markers): "model_score": <integer from 1-5>, // Score for Report A (Model Report) "reference_score": <integer from 1-5>, // Score for Report B (Reference Report) "reasoning": "<Detailed comparison reasoning process, at least 150 words, comprehensively explaining the advantages and disadvantages of both reports and overall differences>", "evidence_model": ["<Specific evidence 1 from model report>", "<Specific evidence 2 from model report>"], "evidence_reference": ["<Specific evidence 1 from reference report>", "<Specific evidence 2 from reference report>"], "suggestions_model": ["<Specific improvement suggestion 1 for model report>", "<Specific improvement suggestion 2>"], "suggestions_reference": ["<Specific improvement suggestion 1 for reference report>", "<Specific improvement suggestion 2>"] VISUAL-TEXT ALIGNMENT [Evaluation Dimension]Visual-Text Semantic Alignment [Definition]Evaluate the formal integrity of visual-text integration, focusing on: (1) Reference clarityâwhether text explicitly references figures (e.g., "as shown in Figure X", "the chart above illustrates"); (2) Transition smoothnessâwhether text naturally leads into figures and provides interpretation afterward; (3) Reading flowâwhether visual-text switching feels natural and integrated into the narrative. [Scoring Rubric]1-5 points 5 points: Seamless visual-text integration with excellent referencing and transitions Text explicitly references each figure with clear pointers. Figures are naturally introduced by preceding text and followed by interpretation/discussion. The reading flow is smoothâfigures feel like integral parts of the narrative, not insertions. Readers never wonder "why is this figure here?" 4 points: Good visual-text integration with clear referencing Most figures have explicit text references. Transitions into and out of figures are generally smooth. Minor instances where a figure appears without clear introduction or follow-up discussion, but overall the integration is coherent. 3 points: Basic visual-text integration with inconsistent referencing Some figures have explicit references, others appear without clear textual connection. Transitions are unevenâsome figures flow naturally, others feel inserted. Readers can follow along but occasionally lose the connection between text and visuals. 2 points: Weak visual-text integration with poor referencing Few explicit figure references. Figures often appear abruptly without introduction or interpretation. Text and visuals feel like separate elements rather than an integrated narrative. Readers must work to understand how figures relate to surrounding text. 1 point: Disconnected visual-text presentation Almost no explicit figure references. Figures appear randomly with no textual connection. Text and visuals are essentially independentâremoving figures would not disrupt text flow (indicating poor integration). Readers cannot understand the visual-text relationship. Multimodal Synergy [Evaluation Dimension]Multimodal Synergy [Definition]Evaluate whether visuals and text work together to create understanding that exceeds what either could achieve alone. Key aspects: (1) Information incrementâwhether figures provide NEW information/perspectives beyond what text states (not just visual repetition of text content); (2) Complementary rolesâwhether text explains concepts while figures show data/relationships/patterns; (3) Synergistic effectâwhether combining text and figures produces 1+1>2 understanding. [Key Distinction] - HIGH synergy: Figure shows data patterns/comparisons that text describes in words âreader gains both conceptual understanding AND visual evidence - LOW synergy: Figure merely visualizes what text already fully explains âfigure is decorative, removing it loses nothing - Ask: "If I remove this figure, would the reader lose important information?" If NO, the figure lacks information increment. [Scoring Rubric]1-5 points 5 points: Excellent synergy with strong information increment Figures provide substantial information beyond textâshowing patterns, comparisons, or relationships that text alone cannot efficiently convey. Text and figures have clear division of labor: text explains "why" and "what it means", figures show "what the data looks like". Removing figures would significantly reduce reader understanding. True 1+1>2 effect. 4 points: Good synergy with meaningful information increment Most figures contribute information beyond text repetition. Text and figures generally complement each other. Some figures may slightly overlap with text content, but overall the combination enhances understanding noticeably. 3 points: Moderate synergy with limited information increment Figures and text have some complementarity, but several figures mainly visualize what text already states. Information increment is inconsistentâsome figures add value, others feel redundant. Removing some figures would not significantly impact understanding. 2 points: Weak synergy, figures largely repeat text Most figures are visual restatements of text content without adding new information or perspectives. Little division of laborâtext and figures say the same things in different formats. Figures feel like illustrations rather than information carriers. 1 point: No synergy, figures are purely decorative Figures provide no information incrementâthey simply convert text statements into visual form. Removing all figures would not reduce information content. Text and figures are redundant rather than complementary. No 1+1>2 effect achieved. Information Organization [Evaluation Dimension]Information Organization Clarity [Definition]Evaluate whether the reportâs structure, layout, and logical connections are clear. This includes: (1) Static structureâwhether hierarchy is clear and complete; (2) Dynamic flowâwhether sections have natural logical progression and smooth transitions. Clear organization can reduce the cognitive cost of visual search and comprehension. [Scoring Rubric]1-5 points 5 points: Perfect structure with excellent logical flow Report structure is complete, hierarchy is clear, sections progress in a natural logical order with smooth transitions between them. Readers can easily follow the reasoning from beginning to end. 4 points: Good structure with reasonable flow Report structure is basically complete, hierarchy is basically clear, sections have reasonable logical order. Transitions between sections are adequate though not always seamless. 3 points: Average structure, weak logical flow Report has basic structure, but logical progression between sections is weak. Some sections feel disconnected or the order seems arbitrary. Readers can understand individual sections but may struggle to see how they connect. 2 points: Messy structure, poor flow Report has some structural elements but lacks clear logical progression. Sections appear randomly ordered, transitions are missing or abrupt. Readers have difficulty following the overall argument. 1 point: No organization, fragmented Report has almost no structure, sections are like fragments randomly pieced together with no logical connection. Readers cannot understand the overall framework or how parts relate. Content Depth and Insight [Evaluation Dimension]Content Depth and Insight [Definition]Evaluate whether the report provides appropriate depth across all important aspects of the topic. This dimension assesses: (1) Coverage completenessâwhether all important facets of the topic are addressed (not just some aspects); (2) Depth balanceâwhether analysis depth is evenly distributed (not deep on some parts while shallow on others); (3) Analytical qualityâwhether the report provides mechanism explanations and causal reasoning, not just facts. [Key Principle] A well-planned report should comprehensively cover the topic with balanced depth across sections. Signs of poor planning include: some sections with rich analysis while others are superficial; important aspects of the topic missing entirely; depth that doesnât match section importance. [Important Clarification] - Depth ̸=Length: A long report with only surface-level facts is NOT deep; a concise report with insightful analysis IS deep - Focus on analytical quality: mechanism explanations, causal reasoning, and meaningful insightsânot word count or section length [Scoring Rubric]1-5 points 5 points: Comprehensive coverage with balanced, high-quality depth Report covers all important aspects of the topic thoroughly. Depth is well-balanced across sectionsâno section feels significantly more superficial or detailed than others relative to its importance. Each section provides meaningful analysis with mechanism explanations and causal reasoning. Readers gain complete understanding of the topic. 4 points: Good coverage with mostly balanced depth Report covers most important aspects with good analytical depth. Depth distribution is reasonable, though minor imbalances exist (e.g., one section slightly more detailed than necessary, another slightly thin). Overall, readers get a solid understanding of the topic. 3 points: Incomplete coverage or unbalanced depth Report has noticeable gaps: either some important aspects of the topic are missing, OR depth is clearly unbalanced (some sections have rich analysis while others are superficial lists). Readers understand parts of the topic well but lack insight into other parts. 2 points: Poor coverage or severely unbalanced depth Report has significant coverage gapsâmultiple important aspects are missing or barely touched. OR depth is severely unbalanced: detailed analysis on minor points while core aspects receive only surface treatment. Readers get fragmented, incomplete understanding. 1 point: Minimal coverage, shallow throughout Report barely covers the topicâmost important aspects are missing. What is covered lacks analytical depth (just facts, no mechanism explanations). Readers cannot form meaningful understanding of the topic. Content Relevance and Adaptation [Evaluation Dimension]Content Relevance and Adaptation [Definition]Evaluate whether the reportâs content is relevant to its stated topic and appropriately structured as a comprehensive report. This dimension assesses: (1) Topic relevanceâwhether all substantive content relates to the reportâs subject matter; (2) Appropriate depthâwhether the report provides sufficient context, background, and analysis expected of a quality report; (3) Non-redundancyâwhether information is presented without excessive repetition across sections. [Important Clarification] - Background sections, methodology explanations, data source descriptions, and contextual information are LEGITIMATE parts of a quality reportâthey should NOT be penalized as "unnecessary content" - Only penalize content that is truly OFF-TOPIC (unrelated to the subject) or EXCESSIVELY REPETITIVE (same points repeated verbatim multiple times) - Meta-elements like citations, acknowledgments, and licensing notices are standard academic/journalistic conventions and should be IGNORED in this evaluation (neither rewarded nor penalized) [Scoring Rubric]1-5 points 5 points: Highly relevant and well-structured report All substantive content directly relates to the report topic. Background, analysis, and conclusions form a coherent whole. No off-topic digressions, no excessive repetition. The report covers the topic comprehensively without wandering. 4 points: Mostly relevant with minor issues Report content is well-aligned with the topic. May have minor digressions or slight repetition, but these do not detract significantly from the overall coherence and relevance. 3 points: Moderately relevant with noticeable issues Report addresses the topic but includes some off-topic sections or noticeable repetition of the same points across different sections. The core content is relevant but diluted by tangential material. 2 points: Poorly focused on the topic Report has significant relevance problems: substantial off-topic content, major digressions from the subject matter, or excessive repetition that makes the report feel padded. Readers struggle to extract the relevant information. 1 point: Largely irrelevant or incoherent Report barely addresses its stated topic. Dominated by off-topic content or so repetitive that little new information is conveyed. The report fails to deliver on its subject matter.