Paper deep dive

Retrieval as Generation: A Unified Framework with Self-Triggered Information Planning

Bo Li, Mingda Wang, Gexiang Fang, Shikun Zhang, Wei Ye

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 77

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%

Last extracted: 4/14/2026, 2:33:59 AM

Summary

GRIP (Generation-guided Retrieval with Information Planning) is a unified framework for Retrieval-Augmented Generation (RAG) that internalizes retrieval control into token-level decoding. By using explicit control tokens ([RETRIEVE], [ANSWER], [INTERMEDIARY], [SOLVED]), the model performs self-triggered information planning, enabling dynamic, multi-step retrieval and reasoning within a single autoregressive trajectory without external controllers.

Entities (4)

GRIP · framework · 100%LLaMA3-8B · model · 95%Self-Triggered Information Planning · methodology · 95%DAPO · algorithm · 90%

Relation Signals (3)

GRIP → uses → Self-Triggered Information Planning

confidence 100% · Central to GRIP is Self-Triggered Information Planning

GRIP → implementedon → LLaMA3-8B

confidence 95% · We fine-tune GRIP on 40,000 structured samples using LLaMA3-8B

GRIP → optimizedby → DAPO

confidence 95% · The model is optimized using DAPO (Yu et al., 2025)

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:We revisit retrieval-augmented generation (RAG) by embedding retrieval control directly into generation. Instead of treating retrieval as an external intervention, we express retrieval decisions within token-level decoding, enabling end-to-end coordination without additional controllers or classifiers. Under the paradigm of Retrieval as Generation, we propose \textbf{GRIP} (\textbf{G}eneration-guided \textbf{R}etrieval with \textbf{I}nformation \textbf{P}lanning), a unified framework in which the model regulates retrieval behavior through control-token emission. Central to GRIP is \textit{Self-Triggered Information Planning}, which allows the model to decide when to retrieve, how to reformulate queries, and when to terminate, all within a single autoregressive trajectory. This design tightly couples retrieval and reasoning and supports dynamic multi-step inference with on-the-fly evidence integration. To supervise these behaviors, we construct a structured training set covering answerable, partially answerable, and multi-hop queries, each aligned with specific token patterns. Experiments on five QA benchmarks show that GRIP surpasses strong RAG baselines and is competitive with GPT-4o while using substantially fewer parameters.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

76,584 characters extracted from source content.

Expand or collapse full text

Retrieval as Generation: A Unified Framework with Self-Triggered Information Planning Bo Li 1 , Mingda Wang 2 , Gexiang Fang 1 , Shikun Zhang 1 , Wei Ye 1 * 1 National Engineering Research Center for Software Engineering, Peking University 2 School of Health Sciences and Biomedical Engineering, Hebei University of Technology deepblue.lb@gmail.com, wye@pku.edu.cn § WisdomShell/GRIP GRIP Project* HuggingFace Model Abstract We revisit retrieval-augmented generation (RAG) by embedding retrieval control directly into generation. Instead of treating retrieval as an external intervention, we express re- trieval decisions within token-level decoding, enabling end-to-end coordination without ad- ditional controllers or classifiers. Under the paradigm of Retrieval as Generation, we pro- pose GRIP (Generation-guided Retrieval with Information Planning), a unified framework in which the model regulates retrieval behav- ior through control-token emission. Central to GRIP is Self-Triggered Information Planning, which allows the model to decide when to re- trieve, how to reformulate queries, and when to terminate, all within a single autoregressive trajectory. This design tightly couples retrieval and reasoning and supports dynamic multi-step inference with on-the-fly evidence integration. To supervise these behaviors, we construct a structured training set covering answerable, par- tially answerable, and multi-hop queries, each aligned with specific token patterns. Experi- ments on five QA benchmarks show that GRIP surpasses strong RAG baselines and is com- petitive with GPT-4o while using substantially fewer parameters. 1 Introduction Retrieval-Augmented Generation (RAG) improves LLMs by grounding generation on external evi- dence, boosting open-domain QA, fact verification, and multi-hop reasoning (Gao et al., 2023; Kand- pal et al., 2023; Xu et al., 2024; Luo et al., 2024; Yang et al., 2025b). However, most RAG systems keep retrieval external and one-shot: they retrieve documents from the initial query and then generate from a fixed context (Yu et al., 2023; Shi et al., 2023; Wang et al., 2023), an assumption that can fail when information needs emerge gradually dur- * Corresponding author ing step-by-step reasoning or when queries contain ambiguity and hidden dependencies. To address this limitation, prior work explores re- trieval control to improve timing and evidence use. Training-free methods (e.g., DRAGIN (Su et al., 2024), FLARE (Jiang et al., 2023)) trigger retrieval via heuristic uncertainty signals, while training- based approaches learn retrieval-related behaviors with supervision, spanning both single-step control (e.g., INFO-RAG (Xu et al., 2024), Self-RAG (Asai et al., 2024), GainRAG (Jiang et al., 2025)) and agentic deep-search systems that iteratively gener- ate queries (e.g., R1-Searcher (Song et al., 2025), ZeroSearch (Sun et al., 2025)). However, many of these methods still rely on auxiliary controllers or multi-stage procedures outside token-level genera- tion, so retrieval timing, query reformulation, and stopping are not represented as explicit, trainable actions within a single decoding trajectory. This separation can also make it harder to attribute er- rors to specific decisions (e.g., retrieving too early vs. stopping too late) and to learn a consistent pol- icy that couples retrieval with the model’s evolving intermediate states. Rather than relying on external modules to make discrete retrieval decisions, we embed retrieval be- havior directly into the model’s generative pol- icy. Under the paradigm of Retrieval as Gen- eration, we propose GRIP (Generation-guided Retrieval with Information Planning), a unified framework where retrieval is controlled by the same token-level decoding process as language generation 1 . GRIP regulates retrieval via explicit control tokens, including[RETRIEVE],[ANSWER], [INTERMEDIARY], and[SOLVED]. At its core, self- triggered information planning allows the model 1 Here, ‘Retrieval as Generation’ does not mean that the retriever is internalized into the LLM. The retriever remains ex- ternal; GRIP generates retrieval-control actions (e.g., retrieval triggering, query reformulation, and termination decisions) as part of the decoding trajectory, and the retrieved documents are then supplied back as context. arXiv:2604.11407v1 [cs.CL] 13 Apr 2026 to decide when to retrieve, formulate follow-up queries conditioned on the evolving reasoning con- text, and determine when to stop and finalize the answer, all within a single autoregressive trajectory. To support these behaviors, we introduce a struc- tured training paradigm that supervises four answer- ability types with distinct control-token patterns. This supervision teaches the model to judge infor- mation sufficiency, trigger retrieval when needed, and terminate once the question is resolved. Un- like heuristic triggering, GRIP learns retrieval as a token-driven and context-sensitive capability. By integrating retrieval planning into generation, GRIP treats retrieval as an integral part of reasoning, enabling multi-step inference, adaptive retrieval depth, and robust generalization across tasks. We conduct experiments on five knowledge- intensive QA benchmarks using multiple metrics. GRIP outperforms strong baselines and is competi- tive with proprietary API models. We further pro- vide behavioral analyses of retrieval timing, query generation quality, and termination control, com- plemented by case studies and visualizations. In summary, our main contributions are as follows: •We propose GRIP, a unified Retrieval as Gen- eration framework that internalizes retrieval into token-level decoding via explicit control tokens, enabling retrieval planning, query re- formulation, and termination within a single autoregressive trajectory. • We extensively evaluate GRIP on five QA benchmarks and provide a mechanistic anal- ysis of retrieval-as-generation behaviors. Be- yond consistent gains over strong baselines, our studies show that GRIP learns task-aware retrieval depth and reliable stopping, yield- ing more adaptive and controllable retrieval trajectories across datasets. 2 Generation-guided Retrieval with Information Planning 2.1Token-Level Control for Unified Retrieval and Generation We propose GRIP, a Retrieval-as-Generation framework that internalizes retrieval control into token-level decoding. By emitting a small set of control tokens, the model decides when to retrieve, how to reformulate queries, and when to terminate. At the core is self-triggered information planning, which assesses information sufficiency and exe- cutes multi-step retrieval. To realize this control in a unified and interpretable manner, GRIP aug- ments the model’s output space with a small set of control tokens, allowing retrieval and termination decisions to be made directly during decoding. GRIP introduces four tokens:[RETRIEVE](re- quest external evidence),[INTERMEDIARY](emit a partial state),[ANSWER](start the final response), and[SOLVED](terminate).They form struc- tured patterns that define two main branches: [INTERMEDIARY]→[RETRIEVE]for continuing ev- idence acquisition, and[ANSWER]→[SOLVED]for completion, enabling multi-step and recursive re- trieval when needed. This minimal interface keeps control within au- toregressive generation, enabling end-to-end learn- ing without external controllers. 2.2 Self-Triggered Information Planning GRIP adopts a language-native mechanism called self-triggered information planning, which enables the model to autonomously decide when and how to retrieve information during generation. This pro- cess unfolds autoregressively, driven by the emis- sion of control tokens. GRIP organizes its planning behavior into the following loop: 1. Initial Decision: Given an original query, the model first determines whether its internal knowl- edge suffices. • If so, it emits[ANSWER]followed by a fi- nal prediction, and then terminates with [SOLVED]. •If not, it emits[INTERMEDIARY]to provide partial knowledge or reasoning, followed by [RETRIEVE]and generates a new query for external retrieval. 2. Retrieval and New Query Generation: The model uses the original or generated query to re- trieve relevant passages, which are incorporated as augmented input. It then reassesses whether the available information is sufficient: •If the evidence is sufficient, the model emits [ANSWER]followed by a final prediction and concludes with [SOLVED]. •Otherwise, it generates[INTERMEDIARY]ex- planation and emits a follow-up[RETRIEVE] token with a new query to continue the loop. Wikipedia QA Dataset BM25GPT-4o-mini Type-α Direct Answer Type-β Retrieval Needed Type-γ Multi-hop Planning Type-θ Answer Completion EM & CoverEM Filtering LLM LLM Supervised Finetuning Rule-based Reinforcement Learning LLM ctrl r Answer Quality Behavior Control GRIP "Question": Original Query "Intermediate Answer": " " "Retrieved Context": " " Original Input "Output": [ANSWER] Answer. [SOLVED] Direct Answer Training Inference Token-Controlled Decoding Signals [ANSWER] Start Final Answer [RETRIEVE] Trigger External Retrieval [SOLVED] Final Termination [INTERMEDIARY] Partial Response Self-Triggered Information Planning Learning Token-Controlled Retrieval Behavior via Structured Supervision Is Knowledge Sufficient "Output": [INTERMEDIARY] Partial Answer. [RETRIEVE] Original Query. Retrieval Needed Is Knowledge Sufficient "Output": [ANSWER] Answer. [SOLVED] Answer Completion Multi-hop Planning "Output": [INTERMEDIARY] New Partial Answer. [RETRIEVE] New Query. Terminate After N Steps Retrieved Context Added ans r Figure 1: Overview of the GRIP framework. The top half illustrates the training process: four behavior types are constructed for supervised fine-tuning, followed by reinforcement learning to improve answer quality and retrieval control. The bottom half depicts self-triggered information planning, where GRIP dynamically emits control tokens to decide whether to retrieve, generate new queries, or terminate reasoning, all within a unified decoding trajectory. Notably, GRIP distinguishes between initial and follow-up retrievals: 1) In the initial decision step, the[RETRIEVE]token is followed by the original query, as the model lacks any retrieved context. 2) In subsequent planning steps, the model conditions on its previously generated intermediate response and the original query to synthesize a new follow- up query tailored to unresolved aspects of the task. 3. Multi-hop Planning: This alternation con- tinues recursively:each[INTERMEDIARY]re- flects the model’s current understanding, and each [RETRIEVE]introduces a targeted query to acquire additional knowledge. 4. Termination: The retrieval-reasoning loop concludes when the model emits[ANSWER]fol- lowed by[SOLVED], indicating that sufficient infor- mation has been gathered to produce a final answer. To ensure efficiency and behavioral stability, GRIP uses a decoding-time retrieval budget, which can be adjusted at inference time and is set to three re- trieval rounds by default. If this limit is reached, the model is required to finalize its response: it emits [ANSWER]and generates the final answer based on the currently available information. This planning mechanism is fully internal to the model and operates within a unified decoding flow. No confidence thresholds, external classifiers, or prompt chaining are needed. Through supervised training on structured examples, the model learns to coordinate retrieval and reasoning as a compo- sitional skill. Pseudocode for the GRIP decoding process is provided in Appendix M. 2.3 Learning Token-Controlled Retrieval Behavior via Structured Supervision To enable GRIP to learn retrieval as a generation- internal process, we formulate a structured super- vision paradigm that aligns specific control token patterns with different retrieval behaviors. We or- ganize training examples into four structured su- pervision scenarios, each reflecting a distinct be- havioral trajectory. This token-controlled supervi- sion allows GRIP to acquire compositional retrieval strategies, such as when to retrieve, how to formu- late a new query, and when to terminate reasoning. We illustrate the four supervision types and their corresponding token structures in Figure 2. Type-αsamples represent directly answerable queries. These are queries that can be answered cor- rectly using the model’s internal knowledge alone, without any external context. We identify such instances by running the backbone model (e.g., LLaMA3-8B-instruct) and filtering cases that con- sistently yield exact-match (EM) answers across multiple decoding attempts. The model is trained to generate[ANSWER]followed by the final answer, and conclude with [SOLVED]. Type-α (Direct Answer) "Question": Original Query "Intermediate Answer": " " "Retrieved Context": " " "Output": "[ANSWER] Answer. [SOLVED]" Type-β (Retrieval Needed) "Question": Original Query "Intermediate Answer": " " "Retrieved Context": " " "Output": "[INTERMEDIARY] Partial Answer. [RETRIEVE] Original Query." Type-γ (Multi-hop Planning) "Question": Original Query "Intermediate Answer": "Partial Answer" "Retrieved Context": "Retrieved Passages" "Output": "[INTERMEDIARY] New Partial Answer. [RETRIEVE] New Query." Type-θ (Answer Completion) "Question": Original Query "Intermediate Answer": "Partial Answer" "Retrieved Context": "Retrieved Passages" "Output": "[ANSWER] Answer. [SOLVED]" Figure 2: Four types of structured training samples in GRIP, each aligned with specific token patterns for controlling retrieval behavior. This design enables the model to learn when to retrieve, plan new queries, or finalize answers via language-native supervision. Type-βsamples involve cases where the model provides partially correct or noisy answers that con- tain the gold answer but lack clarity. These cases are detected using a relaxed coverage-based EM metric, defined as responses that contain the target answer but fail to fully resolve the query. Here, the model is expected to emit[INTERMEDIARY] to reflect its current partial knowledge, followed by[RETRIEVE]and a retrieval query. This setup teaches the model to recognize uncertainty and trig- ger retrieval appropriately. Type-γsamples simulate multi-hop or com- plex knowledge cases where neither the model nor basic retrieval suffices.We identify such samples by detecting failures to answer correctly even with BM25-based retrieval, and then rely on GPT-4o-mini to generate an improved follow- up query focused on missing aspects.The model is trained to generate an intermediate answer ([INTERMEDIARY]) and a refined query ([RETRIEVE]), and to iterate through multiple plan- ning steps before concluding with[ANSWER]and [SOLVED]. Type-θsamples are cases where both the model and retriever can find passages that cover the an- swer superficially, but the correct response requires compressing, selecting, or synthesizing salient de- tails. We use CoverEM to label these samples and train the model to refine retrieved content through reasoning. This teaches the model to emit token transitions like[INTERMEDIARY] → [RETRIEVE] → [ANSWER], modeling selective information us- age in noisy contexts. Token supervision is applied using standard teacher forcing over these structured sequences. By associating each sample type with a precise token trajectory, the model learns to internalize retrieval control within its generative policy. Representative examples are provided in Appendix D. 2.4 Training Objective and Optimization GRIP is trained in two stages: (1) supervised fine-tuning on structured behavioral examples to learn retrieval patterns, and (2) rule-based rein- forcement learning (RL) guided by two reward signals. Together, these stages enable the model to acquire retrieval planning skills that are both interpretable and adaptive. Supervised Fine-Tuning. The initial phase of GRIP training employs supervised learning to ex- pose the model to diverse token-driven retrieval behaviors. Each training instance is labeled with a structured target sequence containing both rea- soning content and control tokens (e.g.,[ANSWER], [RETRIEVE]), reflecting the semantic structure of the corresponding query type. The objective is to minimize the standard autoregressive cross-entropy loss over the entire target sequence: L SFT =− T X t=1 logP θ (y t | y <t ,x) wherexis the input query, andy 1 ,...,y T is the ground-truth sequence containing both natural lan- guage and special control tokens. Reinforcement Fine-Tuning. While SFT pro- vides strong initial behavior modeling, we observe that the model tends to over-trigger retrieval ac- tions, often emitting multiple[RETRIEVE]tokens even when unnecessary. To address this, we apply a RL phase based on rule-based reward optimization framework. We define two complementary reward signals: 1) Answer Fidelity Rewardr ans . Measures the similarity between the model’s generated an- swer and the ground-truth reference using BLEU score (Chang et al., 2025), encouraging accurate and semantically faithful responses. 2) Control Accuracy Rewardr ctrl . Rewards the correct emission of control token patterns and pe- nalizes incorrect or missing tokens, thereby sharp- ening the model’s understanding of retrieval control logic. The final reward is computed as: R = r ans + r ctrl The model is optimized using DAPO (Yu et al., 2025), encouraging answer-faithful and token- accurate behaviors that promote a more balanced retrieval strategy. Overall, GRIP is designed to unify retrieval plan- ning and language generation under a single au- toregressive framework. Compared with closely related search-based methods, its advantage comes from two aspects: (1) retrieval planning is inter- nalized as token-level control within generation itself, rather than implemented through external search-action loops; and (2) multi-step retrieval behavior is learned via one-step decision opti- mization, rather than long-horizon search-policy optimization. This makes GRIP both conceptually simple and practically stable, while still enabling adaptive retrieval depth and controllable stopping behavior at inference time. 3 Experimental Setup and Main Results 3.1 Datasets and Evaluation Metrics Datasets.We evaluate GRIP on five widely used benchmarks for open-domain and multi-hop ques- tion answering: HotpotQA (Ho et al., 2020), PopQA (Mallen et al., 2022), Natural Questions (NQ) (Kwiatkowski et al., 2019), WebQuestions (WebQ) (Berant et al., 2013), and TriviaQA (Joshi et al., 2017). These datasets cover diverse query types, including factual recall, contextual under- standing, and multi-step reasoning, providing a comprehensive testbed for retrieval-augmented gen- eration. We use the official evaluation splits and verify that there is no overlap between evaluation questions and any training data used in our experi- ments. See Appendix A for more detail. Evaluation Metrics. We report three primary QA metrics: Exact Match (EM) (exact normal- ized string match), ROUGE (average of ROUGE- 1/2/L), and F1 (token-level F1 between prediction and reference). For compact reporting, we also use Avg.Score, defined as the unweighted mean of all reported metric values across all evaluated datasets (i.e., averaging over datasets and metrics with equal weight). We additionally report Cov- erEM for behavioral analysis; its definition and results are provided in Appendix C. 3.2 Experimental Details Structured Training Data Construction. We construct structured training data based on the train- ing sets of TriviaQA and NQ, each paired with three retrieved passages. The same top-3 retrieval setting is applied during inference for consistency. Specifically, we construct 40,000 structured train- ing instances for the supervised fine-tuning stage, covering all four behavior types described in Sec- tion 2.3. For the RL stage, we generate an addi- tional 5,000 samples with the same distribution across behavior types. The prompts used to con- struct structured data are provided in Appendix K. Additional settings for control tokens are described in Appendix E. Unless otherwise specified, we set the maximum retrieval steps to three, and we ana- lyze larger test-time budgets in § 4.4. Supervised Fine-tuning. We fine-tune GRIP on 40,000 structured samples using LLaMA3- 8B (Grattafiori et al., 2024) as the backbone with full-parameter training. The model is trained for 8 epochs with a micro-batch size of 4 per GPU (total batch size 32 across 8 A800 GPUs). We use the AdamW with a learning rate of1× 10 −6 . A cosine learning rate schedule is applied with 10% warm-up. Input sequences are truncated at 8192 tokens. Due to space limitations, results using the Qwen2.5-7B-Instruct (Yang et al., 2024, 2025a) are provided in the Appendix P. ReinforcementLearning. Weadoptthe DAPO (Yu et al., 2025) algorithm, an improved variant of GRPO (Shao et al., 2024), to fine-tune GRIP on 5,000 structured samples. Two task- specific rewards are used: (1) Answer Fidelity, computed via BLEU score against the reference answer, more details about the reward selection ablation can be found in the Appendix F; and (2) Control Accuracy, which assigns 0.5 points for correctly emitted control tokens. We train the model for only 1 epoch with a learning rate of 1× 10 −7 , which is sufficient to stabilize retrieval behaviors under the DAPO framework. Please refer to Appendix O for more details. 3.3 Comparison Models We compare GRIP with two groups of baselines. Training-free methods include direct prompting (Instruct), one-shot retrieval (Single RAG), pro- prietary API models (GPT-3.5 Turbo, GPT-4o) as references, and dynamic retrieval methods FLARE (Jiang et al., 2023), DRAGIN (Su et al., 2024) and ETC (Li et al., 2026) (8-shot ICL follow- ing their original setups). Training-based methods include a matched-data SFT baseline (SFT-RAG) and state-of-the-art frameworks Self-RAG (Asai HotpotQAPopQANQWebQTriviaQAAvg. ScoreEM ROUGE F1EM ROUGE F1EM ROUGE F1EM ROUGE F1EM ROUGE F1 Training-free Method Instruct17.221.625.917.419.623.214.423.820.314.825.129.446.245.355.126.6 GPT-3.5 Turbo26.232.438.229.131.135.320.734.330.015.927.431.955.657.869.735.7 GPT-4o33.240.247.030.632.339.926.542.728.323.532.337.065.764.378.241.4 Single RAG26.131.637.222.826.631.119.328.424.814.022.726.646.447.056.830.8 FLARE23.227.932.814.316.018.414.722.421.324.230.334.748.648.556.428.9 DRAGIN27.932.638.715.516.819.823.932.828.525.231.535.755.353.664.633.5 ETC32.537.744.230.532.537.520.926.730.718.926.630.452.952.163.035.8 Training-based Method SFT-RAG20.324.128.629.425.230.420.817.921.318.918.323.150.124.857.227.4 Self-RAG19.623.826.718.122.322.815.722.424.016.426.527.450.247.357.528.0 INFO-RAG 19.923.726.918.322.623.017.222.924.918.127.728.950.847.858.128.7 RobustRAG27.631.837.529.727.732.426.425.129.221.525.029.148.847.757.933.2 GainRAG31.435.641.830.133.338.122.927.932.216.524.528.950.349.159.234.8 R1-Searcher 26.029.134.941.635.241.325.824.928.721.826.130.656.053.364.936.0 RetRobust29.634.940.934.135.140.424.229.033.821.827.431.753.650.961.936.6 InsturctRAG31.236.842.333.135.740.329.529.533.619.526.831.451.352.262.537.0 GRIP(ours)33.037.644.138.637.538.432.135.832.031.439.334.657.955.967.441.0 +4.0 w/o RL 31.636.643.038.137.137.632.636.132.732.039.935.157.055.266.840.7 Table 1: Main results on five QA benchmarks. We report EM, ROUGE (average of ROUGE-1/2/L), and F1; Avg.Score is the unweighted mean over all datasets and metrics. All open-source baselines are reproduced using official implementations and released checkpoints under the same evaluation setting, and results are averaged over three random seeds. Results on a domain-specific dataset are provided in Appendix B. et al., 2024), INFO-RAG (Xu et al., 2024), Ro- bustRAG (Fang et al., 2024),RetRobust (Yoran et al.), InstructRAG (Wei et al., 2024), Robus- tRAG (Fang et al., 2024), GainRAG (Jiang et al., 2025), and R1-Searcher (Song et al., 2025). To ensure a fair comparison, we reproduce all open-source baselines with official code/check- points when available and evaluate all models un- der the same protocol: LLaMA3-8B backbone, Wikipedia corpus, BM25 retriever with top-3 pas- sages per query. For methods without explicit an- swer spans (e.g., INFO-RAG, RobustRAG, Self- RAG), we apply minimal output standardization for consistent scoring. We ablate retrievers (DPR and BM25+DPR) in Appendix L and use BM25 by default for its performance–efficiency trade-off. 3.4 Main Results Table 1 presents a comprehensive comparison be- tween GRIP and a broad range of RAG methods. GRIP consistently achieves the best overall perfor- mance among open-source systems, outperforming strong training-based baselines such as GainRAG and RobustRAG across all five benchmarks, and also surpassing recent deep-search baselines such as R1-Searcher in Avg.Score. By internalizing retrieval decisions into the autoregressive gener- ation process via token-level control, GRIP elimi- nates the need for external heuristics or controllers. Despite using a smaller backbone (LLaMA3-8B), GRIP reaches an Avg.Score of 41.0, close to GPT- 4o, demonstrating strong effectiveness under a lightweight setting. Gains are most pronounced on multi-hop or com- positional benchmarks (e.g., HotpotQA and Trivi- aQA), where dynamic evidence acquisition is cru- cial. Beyond exact matching, GRIP also yields con- sistently higher ROUGE and F1, indicating better answer completeness and alignment. Overall, these results support GRIP as an efficient and generaliz- able framework for controlled retrieval-augmented generation across diverse QA scenarios. 4 Analysis 4.1 Retrieval-Depth Adaptivity A central benefit of token-level planning is bidi- rectional control over retrieval depth: GRIP sup- presses retrieval when the current decoding state is already sufficient, yet deepens retrieval when evi- dence is missing. To make this behavior concrete, we study two complementary subsets. First, we consider cases where DRAGIN/Gain- RAG perform at least one retrieval while GRIP an- swers directly; Table 2 shows that GRIP achieves NQ(54.3%)WebQ(24.3%)Avg. Score EM ROUGE F1EM ROUGE F1 DRAGIN30.333.939.342.548.452.941.2 GainRAG31.634.439.633.040.745.637.4 GRIP 31.134.039.844.950.755.642.7 Table 2: Subset where DRAGIN/GainRAG retrieve at least once while GRIP answers without retrieval. Per- centages denote the subset proportion per dataset. NQ(23.1%)WebQ(43.8%)Avg. Score EM ROUGE F1EM ROUGE F1 DRAGIN11.416.719.018.623.828.319.6 GainRAG8.814.817.411.118.422.415.5 GRIP15.420.323.419.924.628.922.1 Table 3: Subset where GRIP performs two retrievals while DRAGIN/GainRAG retrieve at most once. Per- centages denote the subset proportion per dataset. higher accuracy on this subset despite using fewer retrieval calls, indicating it can avoid redundant retrieval rather than following a fixed schedule. Second, we analyze cases where GRIP performs two retrievals while baselines retrieve at most once; Table 3 shows consistent gains on these harder ex- amples, suggesting GRIP can detect evidence insuf- ficiency and extend retrieval accordingly. Together, these results support that GRIP learns context- sensitive retrieval depth control from structured token supervision. 4.2 Adaptive Retrieval Across Tasks Table 4 reports the mean number of retrieval calls per example. GRIP exhibits clear task-aware adap- tivity in retrieval depth: it retrieves more frequently on HotpotQA (1.44) and PopQA (1.58), where multi-hop reasoning or long-tail entities often re- quire external evidence (Mallen et al., 2022; Asai et al., 2024), while triggering substantially fewer re- trievals on NQ (0.76), where many questions can be answered from the model’s parametric knowledge. Compared with training-based baselines that fol- low nearly fixed retrieval schedules (e.g., GainRAG always retrieves once), GRIP adjusts retrieval fre- quency across datasets without external controllers. Moreover, R1-Searcher performs markedly more retrievals (4.67–5.75 across datasets; 5.12 on aver- age), indicating a substantially higher retrieval cost despite its weaker end performance in our setting. In contrast, GRIP achieves strong accuracy with a much lower retrieval budget. Hotpot PopQA NQ WebQ Trivia Avg. Count DRAGIN1.131.261.06 1.021.081.11 GainRAG1.01.01.01.01.01.0 R1-Searcher5.755.345.09 4.674.765.12 GRIP 1.441.580.76 1.151.251.24 w/o RL1.991.980.77 1.261.991.60 Table 4: Mean retrieval count per dataset. GRIP adapts retrieval frequency across tasks, and RL further reduces retrieval calls while preserving this adaptivity. Hot- pot and Trivia denote HotpotQA and TriviaQA, respec- tively. Besides, rule-based RL further refines GRIP’s retrieval policy by consistently reducing redundant calls across all datasets: the overall average drops from 1.60 to 1.24 (∼22.5% reduction), while pre- serving the same task-aware retrieval pattern. A finer-grained breakdown of this behavioral shift is provided in Appendix G. 4.3 Improving Retrieval Quality by Generating New Queries Top-1Top-3 0 5 10 15 20 25 Answer Coverage (%) 8.18% 12.88% 13.12% 20.70% NQ Original Query GRIP-generated Top-1Top-3 Answer Coverage (%) 7.60% 13.18% 18.99% 25.81% WebQ Figure 3: The percentage of samples where the gold answer span is found in the top-1 or top-3 retrieved passages. GRIP-generated queries significantly improve answer presence in retrieved contexts across both NQ and WebQ, indicating more effective retrieval behavior. To evaluate the effectiveness of GRIP’s self- triggered information planning, we analyze the retrieval quality of its generated new queries com- pared to the original ones. Specifically, we measure the proportion of top-1 and top-3 retrieved passages that contain the gold answer. As shown in Figure 3, GRIP significantly improves retrieval coverage on both NQ and WebQ. This improvement does not stem from superficial query rewriting. Instead, it emerges from GRIP’s iterative reasoning process. By leveraging intermediate answers and previously retrieved content, GRIP gradually refines its focus toward unresolved aspects of the question. This progressive formulation yields more targeted new queries that are more likely to retrieve passages con- taining the correct answer. Notably, larger gains on top-1 than top-3 suggest that GRIP improves evi- dence ranking, consistent with query reformulation being coupled to intermediate reasoning states for more efficient retrieval. 4.4 Controllable Retrieval Budget and Depth Extrapolation We vary the maximum allowed retrieval stepsBat inference time and report both the realized retrieval frequency (Avg.Count) and overall Avg.Score (Ta- ble 5). AsBincreases from 3 to 10, Avg.Score rises from 41.0 to 41.8 while Avg.Count grows only from 1.24 to 1.62. Avg.Count stays far be- low the budget cap across all settings, indicating that GRIP does not blindly consume the available budget but retrieves adaptively when the decoding trajectory suggests an information gap, supporting our token-level planning claim. Max B Avg.Count (per dataset)Avg. Score Hotpot PopQA NQ WebQ Trivia Avg 31.441.580.76 1.151.25 1.2441.0 5 1.561.731.01 1.451.31 1.4141.2 71.661.791.12 1.651.35 1.5141.5 101.741.861.22 1.901.39 1.6241.8 Table 5: Effect of the maximum retrieval budgetBat inference time. Hotpot and Trivia denote HotpotQA and TriviaQA, respectively. This adaptive control also exhibits extrapola- tive generalization: although training trajectories contain at most three retrieval steps, GRIP ben- efits from larger test-time budgets (B=5, 7, 10) with consistent gains, suggesting it learns when additional evidence is needed rather than a fixed- step pattern. Extra retrieval is mostly allocated to evidence-uncertain benchmarks (NQ: 0.76→1.22; WebQ: 1.15→1.90), and improvements show di- minishing returns. Overall,Bprovides a practical control parameter to trade accuracy for cost, while GRIP remains adaptive beyond the retrieval depth seen in training. Appendix N shows the distribution of retrieval counts. 4.5 Replacing Proprietary Teachers with Open-Source Supervision To test whether GRIP’s gains rely on propri- etary distillation, we conduct a controlled teacher- replacement study on Type-γsupervision, which is the only place where our data construction uses an external teacher. Specifically, Type-γsamples use TeacherTrainingAvg.Score GPT-4o-mini GRIP (SFT)40.7 GRIP (SFT+RL)41.0 LLaMA3-8B-Instruct GRIP (SFT)38.8 GRIP (SFT+RL)39.0 Qwen3-32B GRIP (SFT)40.2 GRIP (SFT+RL)40.3 Table 6: Teacher replacement results, where only the GPT-labeled subset is re-annotated. GPT-4o-mini to generate improved follow-up re- trieval queries when the backbone model fails. We replace GPT-4o-mini and re-annotate the follow- up queries for all Type-γsamples using two open teachers: LLaMA3-8B-Instruct and Qwen3-32B. All other components are kept identical across set- tings, including the construction and supervision of other samples, the prompts and decoding hyper- parameters for annotation. Table 6 shows that GRIP remains strong under open teachers, indicating that the gains primarily arise from our framework design rather than GPT- specific artifacts. In particular, replacing GPT-4o- mini with Qwen3-32B yields a comparable final performance (40.3 vs. 41.0), suggesting that GRIP can still induce effective dynamic retrieval and query reformulation behaviors without relying on closed-source supervision. Using a backbone-scale teacher (LLaMA3-8B-Instruct) leads to a larger drop, implying that teacher capacity mainly affects the quality of trajectory instantiation rather than the existence of the learned retrieval policy. Overall, teacher outputs serve as interchangeable demon- strations, while the token-controlled supervision is the key factor that turns retrieval into a learnable generation skill. 4.6 General Capability Preservation ModelMMLU (Acc)MBPP (Pass@1) Instruct66.5654.6 SFT-RAG62.9347.8 GRIP65.7353.8 Table 7: General capability evaluation on closed-form non-RAG benchmarks. A potential concern is whether fine-tuning GRIP for retrieval planning may degrade the model’s general-purpose abilities on non-RAG tasks. To directly assess this, we conduct additional eval- Pairwise comparison (GRIP vs.)WinEqualLoss Instruct51.037.511.5 SFT-RAG90.58.51.0 R1-Searcher59.532.58.0 Table 8: Pairwise GPT-4o evaluation on CNN/Daily- Mail summarization. Results are averaged over both output orders to mitigate position bias. uations on two types of tasks that involve no re- trieval: (1) closed-form knowledge and coding benchmarks, including MMLU (Hendrycks et al., 2020) and MBPP (Austin et al., 2021), and (2) an open-ended generation task, CNN/DailyMail sum- marization (Hermann et al., 2015), evaluated by GPT-4o as a pairwise judge. Closed-form benchmarks. Table 7 reports re- sults on MMLU and MBPP. Compared with the base INSTRUCT model, GRIP shows only a mi- nor decrease, dropping by 0.83 points on MMLU and 0.8 points on MBPP. This suggests that learn- ing retrieval-planning behaviors does not substan- tially harm the model’s general knowledge, rea- soning, or coding ability. In contrast, the matched fine-tuning baseline SFT-RAG exhibits noticeably larger degradation on both benchmarks, indicating that GRIP better preserves the backbone model’s general capabilities while still introducing retrieval- oriented behaviors. Open-ended summarization. We further evalu- ate open-ended generation quality on CNN/Daily- Mail summarization, where no retrieval is involved. Following recent practice, we use GPT-4o as a pair- wise judge to compare model outputs. To reduce position bias, each model pair is evaluated in both orders, and we report the order-averaged Win/E- qual/Loss rates in Table 8. GRIP remains compet- itive with the base INSTRUCT model, achieving a 51.0% win rate versus 11.5% loss rate, and is pre- ferred much more often than other fine-tuned or search-oriented baselines such as SFT-RAG and R1-SEARCHER. These results suggest that GRIP preserves general-purpose generation quality while learning retrieval control. Overall, across both closed-form and open- ended non-RAG evaluations, GRIP shows min- imal degradation relative to the base INSTRUCT model, while substantially outperforming retrieval- oriented baselines in preserving general-purpose capabilities. This indicates that integrating retrieval planning into token-level generation does not come at the cost of broadly useful language modeling ability. 5 Related Work Most standard RAG pipelines retrieve documents in a static, one-shot fashion based on the initial query (Gao et al., 2023; Fan et al., 2024; Xiong et al., 2024; Ye et al., 2024; Zhang et al., 2025, 2026). To reduce unnecessary retrieval, several works estimate retrieval necessity at inference time using heuristic uncertainty signals, such as entropy- based filtering or logit-margin thresholds (Jiang et al., 2023; Ram et al., 2023; Su et al., 2024; Li et al., 2025). While lightweight, these approaches rely on hand-crafted criteria rather than learning retrieval behaviors end-to-end. Recent studies make retrieval behavior learnable by introducing retrieval-relevant supervision, e.g., self-reflection/critique signals or contrastive/adver- sarial objectives (Asai et al., 2024; Jiang et al., 2025; Xu et al., 2024; Fang et al., 2024). In parallel, deep search and multi-step retrieval systems further strengthen search-and-reason routines for complex queries (Song et al., 2025; Sun et al., 2025), as well as planning-oriented multi-hop methods (Lee et al., 2024; Jin et al., 2025; Yan and Ling, 2025; Zhu et al., 2026). However, many of these ap- proaches still rely on external controllers or multi- stage pipelines for retrieval decisions, which can complicate integration and reduce transparency. In contrast, GRIP internalizes retrieval as token-level generative actions, realizing timing, query reformu- lation, and termination within a single autoregres- sive trajectory. 6 Conclusion We present GRIP, a unified retrieval-augmented generation framework that integrates retrieval be- havior into the generation process through self- triggered information planning. This design en- ables dynamic control over retrieval without re- lying on external modules. Experimental results across five QA benchmarks confirm that GRIP con- sistently outperforms existing RAG methods and achieves performance comparable to GPT-4o. We hope this work provides a foundation for future efforts in controllable and efficient generation. Limitations Although GRIP learns to regulate retrieval through discrete control tokens, its behavior remains sen- sitive to the design of the retrieval interface, such as the maximum retrieval budgetBand the evi- dence packing strategy (e.g., chunking and top-k). In this work we keep these choices fixed for fair comparison, but more adaptive evidence budget- ing and context structuring could further improve robustness across domains. References Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2024. Self-rag: Learning to re- trieve, generate, and critique through self-reflection. Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and 1 others. 2021. Program synthesis with large language models. arXiv preprint arXiv:2108.07732. Jonathan Berant, Andrew K. Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on freebase from question-answer pairs. In Conference on Em- pirical Methods in Natural Language Processing. Yapei Chang, Yekyung Kim, Michael Krumdick, Amir Zadeh, Chuan Li, Chris Tanner, and Mohit Iyyer. 2025. Bleuberi: Bleu is a surprisingly ef- fective reward for instruction following.ArXiv, abs/2505.11080. Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing Li. 2024. A survey on rag meeting llms: Towards retrieval-augmented large language models. In Pro- ceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 6491– 6501. Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiao- jun Chen, and Ruifeng Xu. 2024. Enhancing noise robustness of retrieval-augmented language models with adaptive adversarial training. In Annual Meeting of the Association for Computational Linguistics. Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. 2023. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language under- standing. arXiv preprint arXiv:2009.03300. Karl Moritz Hermann, Tomas Kocisky, Edward Grefen- stette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. Advances in neural information processing systems, 28. Xanh Ho, A. Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. ArXiv, abs/2011.01060. Yi Jiang, Sendong Zhao, Jianbo Li, Haochun Wang, and Bing Qin. 2025. Gainrag: Preference alignment in retrieval-augmented generation through gain signal synthesis. arXiv preprint arXiv:2505.18710. Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023.Ac- tive retrieval augmented generation. arXiv preprint arXiv:2305.06983. Bowen Jin, Hansi Zeng, Zhenrui Yue, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. ArXiv, abs/2503.09516. Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehen- sion. ArXiv, abs/1705.03551. Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In In- ternational Conference on Machine Learning, pages 15696–15707. PMLR. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Red- field, Michael Collins, Ankur P. Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Ken- ton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc V. Le, and Slav Petrov. 2019. Natu- ral questions: A benchmark for question answering research. Transactions of the Association for Compu- tational Linguistics, 7:453–466. Myeonghwa Lee, Seonho An, and Min-Soo Kim. 2024. Planrag: A plan-then-retrieval augmented generation for generative large language models as decision mak- ers. In North American Chapter of the Association for Computational Linguistics. Bo Li, Tian Tian, Zhenghua Xu, Hao Cheng, Shikun Zhang, and Wei Ye. 2025. Modeling uncertainty trends for timely retrieval in dynamic rag. arXiv preprint arXiv:2511.09980. Bo Li, Tian Tian, Zhenghua Xu, Hao Cheng, Shikun Zhang, and Wei Ye. 2026. Modeling uncertainty trends for timely retrieval in dynamic rag. In Pro- ceedings of the AAAI Conference on Artificial Intelli- gence, volume 40, pages 31527–31535. Kun Luo, Zheng Liu, Shitao Xiao, Tong Zhou, Yubo Chen, Jun Zhao, and Kang Liu. 2024. Landmark embedding: A chunking-free embedding method for retrieval augmented long-context large language mod- els. In Annual Meeting of the Association for Com- putational Linguistics. Alex Troy Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Hannaneh Hajishirzi, and Daniel Khashabi. 2022. When not to trust language models: Investigat- ing effectiveness of parametric and non-parametric memories. In Annual Meeting of the Association for Computational Linguistics. Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. In-context retrieval-augmented lan- guage models. Transactions of the Association for Computational Linguistics, 11:1316–1331. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Jun- Mei Song, Mingchuan Zhang, Y. K. Li, Yu Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. ArXiv, abs/2402.03300. Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen tau Yih. 2023. Replug: Retrieval-augmented black-box language models. In North American Chapter of the Association for Computational Lin- guistics. Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji- Rong Wen. 2025. R1-searcher: Incentivizing the search capability in llms via reinforcement learning. ArXiv, abs/2503.05592. Weihang Su, Yichen Tang, Qingyao Ai, Zhijing Wu, and Yiqun Liu. 2024. Dragin: Dynamic retrieval aug- mented generation based on the real-time informa- tion needs of large language models. arXiv preprint arXiv:2403.10081. Hao Sun, Zile Qiao, Jiayan Guo, Xuanbo Fan, Yingyan Hou, Yong Jiang, Pengjun Xie, Yan Zhang, Fei Huang, and Jingren Zhou. 2025. Zerosearch: Incen- tivize the search capability of llms without searching. ArXiv, abs/2505.04588. Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query expansion with large language models. In Conference on Empirical Methods in Natural Language Processing. Zhepei Wei, Wei-Lin Chen, and Yu Meng. 2024. In- structrag: Instructing retrieval-augmented genera- tion via self-synthesized rationales. arXiv preprint arXiv:2406.13629. Haoyi Xiong, Jiang Bian, Yuchen Li, Xuhong Li, Meng- nan Du, Shuaiqiang Wang, Dawei Yin, and Sumi Helal. 2024. When search engine services meet large language models: visions and challenges. IEEE Transactions on Services Computing. Shicheng Xu, Liang Pang, Mo Yu, Fandong Meng, Huawei Shen, Xueqi Cheng, and Jie Zhou. 2024. Un- supervised information refinement training of large language models for retrieval-augmented generation. ArXiv, abs/2402.18150. Shi-Qi Yan and Zhen-Hua Ling. 2025.Rpo: Re- trieval preference optimization for robust retrieval- augmented generation. ArXiv, abs/2501.13726. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayi- heng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 oth- ers. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jian- hong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, and 22 oth- ers. 2024. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. Yuming Yang, Jiang Zhong, Li Jin, Jingwang Huang, Jingpeng Gao, Qing Liu, Yang Bai, Jingyuan Zhang, Rui Jiang, and Kaiwen Wei. 2025b. Benchmark- ing multimodal rag through a chart-based document question-answering generation framework. Preprint, arXiv:2502.14864. Fuda Ye, Shuangyin Li, Yongqi Zhang, and Lei Chen. 2024. R2ag: Incorporating retrieval information into retrieval augmented generation. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 11584–11596. Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Be- rant. Making retrieval-augmented language models robust to irrelevant context. In ICLR 2024 Workshop on Large Language Model (LLM) Agents. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, and 16 oth- ers. 2025. Dapo: An open-source llm reinforcement learning system at scale. ArXiv, abs/2503.14476. W. Yu, Zhihan Zhang, Zhenwen Liang, Meng Jiang, and Ashish Sabharwal. 2023. Improving language models via plug-and-play retrieval feedback. ArXiv, abs/2305.14002. Qianchi Zhang, Hainan Zhang, Liang Pang, Yongxin Tong, Hongwei Zheng, and Zhiming Zheng. 2025. Less is more: Compact clue selection for efficient retrieval-augmented generation reasoning. arXiv preprint arXiv:2502.11811. Qianchi Zhang, Hainan Zhang, Liang Pang, Hongwei Zheng, and Zhiming Zheng. 2026. Stable-rag: Miti- gating retrieval-permutation-induced hallucinations in retrieval-augmented generation. arXiv preprint arXiv:2601.02993. Wei Zhu, Zhiwen Tang, and Kun Yue. 2026. Sym- phony: Synergistic multi-agent planning with hetero- geneous language model assembly. arXiv preprint arXiv:2601.22623. Appendix A Test Dataset Statistics A.1 Dataset. We evaluate GRIP on five QA benchmarks as the main evaluation suite in the main paper, and ad- ditionally report results on BioASQ in this ap- pendix as a domain-specific benchmark. HotpotQA focuses on multi-hop reasoning. PopQA, NQ, and WebQ are open-domain QA datasets, target- ing factual questions with varying entity popular- ity and coverage. TriviaQA emphasizes reading comprehension from web and Wikipedia sources, featuring complex, compositional questions au- thored by trivia enthusiasts. Finally, BioASQ con- tains biomedical questions that require specialized knowledge and is not included in the main-suite average unless explicitly stated. These datasets col- lectively test the model’s adaptability to different retrieval and reasoning demands. #SamplesType HotpotQA7,405Multi-hop PopQA 14,267Open-domain NQ 3,610Open-domain WebQ 2,032Open-domain TriviaQA11,313Reading Comprehension BioASQ885Domain-specific Table 9: Statistics and types of the test datasets used in our evaluation. The datasets vary in size and reasoning focus, covering multi-hop, open-domain, reading com- prehension, and biomedical QA tasks. A.2 Evaluation Metrics. We report three automatic metrics on all QA bench- marks, where higher is better. For each test in- stance, we compute the metric against the corre- sponding reference answer(s), and then average scores over the evaluation set. Exact Match (EM). EM is a binary metric that equals 1 if the predicted answer exactly matches any reference answer after standard normalization (e.g., lowercasing and trimming extra whitespaces), and 0 otherwise. The final EM is the average over all instances. Token-level F1. F1 measures the token overlap between the prediction and a reference answer. Fol- lowing the standard QA evaluation protocol, we compute precision and recall based on overlap- ping tokens, and the final F1 is averaged across instances. ROUGE. We compute ROUGE-nforn ∈ 1, 2,L between the generated output and the reference text(s). We then reportROUGE = 1 3 (ROUGE-1 + ROUGE-2 + ROUGE-L). B Performance On The Domain-specific Dataset ROUGEF1Avg.Score FLARE30.269.449.8 DRAGIN47.181.264.2 SFT-RAG39.158.648.9 Self-RAG45.968.857.4 INFO-RAG39.459.249.3 RobustRAG53.980.967.4 GainRAG49.374.061.7 GRIP54.884.469.6 Table 10: Performance on the domain-specific BioASQ benchmark (higher is better). Avg.Score is the un- weighted mean of ROUGE and F1. We further evaluate all models on BioASQ, a domain-specific QA benchmark where questions often require specialized biomedical knowledge be- yond the parametric capacity of general-purpose LLMs. As shown in Table 10, GRIP achieves the best overall performance, obtaining the highest ROUGE (54.8) and F1 (84.4), and thus the highest Avg.Score (69.6). Compared with strong training- based baselines such as RobustRAG and GainRAG, GRIP yields more accurate and better-supported answers, suggesting that its self-triggered infor- mation planning can more reliably decide when to retrieve and how to formulate domain-relevant queries under high knowledge uncertainty. C Cover Match Analysis for Answer Stability To further assess semantic correctness beyond exact matching, we report CoverEM scores across all datasets. CoverEM is designed to measure whether the model output covers the reference answer string. After applying the same text normalization as EM, CoverEM equals 1 if the normalized prediction contains the normalized reference answer string as a contiguous substring (for any reference in the reference set), and 0 otherwise. The final CoverEM is averaged across instances. As shown in Table 11, GRIP significantly out- performs all baselines, including powerful models like GPT-4o and GainRAG. This suggests that even when predictions diverge from the gold answer in surface form, GRIP still generates semantically aligned spans with correct content. The consistent gains across datasets suggest that GRIP’s outputs are not only accurate but also robust to surface-level variations. Hotpot PopQA NQ WebQ Trivia Avg. CoverEM GPT-4o38.645.2 48.5 43.682.151.6 DRAGIN31.623.0 38.4 41.568.040.5 GainRAG38.044.4 39.1 36.964.244.5 GRIP45.049.1 41.2 52.474.952.5 Table 11: CoverEM scores across five QA datasets. GRIP achieves the best average CoverEM. Hotpot and Trivia denote HotpotQA and TriviaQA, respectively. D Structured Dataset Samples We now provide several dataset samples to help readers understand the structured training data eas- ily. Figure 4: Type-α: Samples that are directly answerable using the model’s internal knowledge alone, where the model emits[ANSWER]followed by the answer and con- cludes with [SOLVED] without any retrieval. Figure 5: Type-β: Samples where the model produces partial or noisy answers containing the gold answer but lacking clarity, encouraging it to recognize uncertainty and issue a retrieval request through[INTERMEDIARY] followed by [RETRIEVE]. Figure 6: Type-γ: Multi-hop or complex knowledge samples where both parametric memory and basic re- trieval fail, requiring the model to iteratively construct better sub-questions and reason across multiple planning steps before concluding with[ANSWER]and[SOLVED]. E Various Control Tokens Settings In the preliminary stage of GRIP, we explored mul- tiple strategies for defining control tokens used to guide the decoding process. Specifically, we exper- imented with the following approaches: • Using natural-language-style tokens such as [ANSWER]or[RETRIEVE]directly as special control tokens; • Replacing control tokens with reserved but unused tokens from the pretrained model vocabulary, such as[ANSWER]replaced by <|reserved_special_token_51|>; •Introducing four newly defined tokens into the tokenizer vocabulary and initializing them with embeddings of semantically related words.For example,[RETRIEVE]is ini- tialized with the embedding of the word retrieve. Empirical results showed that these strategies re- sulted in only marginal differences in overall model performance. To reduce implementation complex- ity and ensure compatibility with standard decod- ing processes, we selected the first approach for all subsequent experiments. F Ablation on the Reward Metric We replace the answer-fidelity rewardr ans with BLEU, ROUGE, token-F1, or EM while keeping Figure 7: Type-θ: Cases where the retrieved con- tent covers the answer superficially but demands fur- ther reasoning and content compression, training the model to selectively synthesize key information via [INTERMEDIARY]→ [RETRIEVE]→ [ANSWER]transi- tions. Reward metric for r ans Avg.ScoreAvg.Count BLEU (default)41.01.24 ROUGE 40.51.24 F140.81.25 EM40.61.24 Table 12: Reward-metric ablation for RL: we only change the overlap metric used in the answer-fidelity reward r ans ; all other settings are fixed. the RL pipeline unchanged. As shown in Table 12, performance is highly consistent across metrics (within 0.5 Avg.Score), and the average number of retrieval calls is nearly unchanged. This suggests that the RL stage is not sensitive to a specific over- lap definition and primarily provides fine-grained shaping of answer quality. The near-identical Avg.Count indicates that the control reward dom- inates the retrieval policy, whiler ans mainly af- fects answer wording rather than when to retrieve. We use BLEU by default, as it offers a smoother graded signal than strict EM while achieving the best Avg.Score under essentially the same retrieval budget. G Behavioral Shift after Rule-based RL We analyze GRIP’s retrieval behavior before and after reinforcement learning (GRIP (w/o RL) vs. GRIP) on two representative datasets: NQ and WebQ. As shown in Figure 8, Rule-based RL significantly reduces unnecessary second retrieval steps, e.g., the proportion of Retrieve=2 on WebQ GRIP(w/o RL)GRIP 0 20 40 60 80 100 Percentage (%) 57.8% 60.9% 7.6% 23.2% 34.6% 15.9% NQ GRIP(w/o RL)GRIP 33.8% 30.0% 6.6% 37.5% 59.6% 32.5% WebQ Retrieve Times Retrieve=0 Retrieve=1 Retrieve>=2 Figure 8: Comparison of retrieval behavior distribu- tions before and after reinforcement learning on NQ and WebQ. drops from 59.6% to 32.5%, while Retrieve=1 rises markedly from 6.6% to 37.5%. This shift indicates that GRIP learns to make earlier stopping deci- sions once sufficient context is available, thereby avoiding over-retrieval. On NQ, where many ques- tions can be answered directly, GRIP maintains a high Retrieve=0 rate (from 57.8% to 60.9%), con- firming its ability to preserve efficient strategies on easier tasks. Overall, these results highlight GRIP’s capacity to adapt its retrieval depth based on the complexity of individual questions, demonstrating more interpretable and context-aware control over retrieval behaviors. Overall, RL mainly converts many “Retrieve=2” cases into “Retrieve=1” rather than increasing “Re- trieve=0”, suggesting it improves early stopping once sufficient evidence is obtained. H Case Studies of GRIP’s Retrieval Planning Behavior To better illustrate the internal behavior of GRIP during retrieval-augmented generation, we present a series of qualitative case studies that highlight how the model dynamically plans, regulates, and adapts its retrieval strategy. These examples cover representative scenarios including retrieval under information insufficiency, multi-step query refine- ment, and generation under explicit control sig- nals. Each case demonstrates a different aspect of GRIP’s token-level decision-making process, pro- viding deeper insights into its reasoning trajectory, interpretability, and interaction between generation and retrieval. Case 1: Fallback to Internal Knowledge When GRIP Cannot Answer To understand how GRIP behaves when external in- formation is insufficient or retrieval is delayed, we examine a case where the model fails to answer the original question. Instead of hallucinating a con- fident response, GRIP generates partial reasoning steps based on its internal knowledge, revealing its awareness of information gaps. This demon- strates GRIP’s ability to initiate cautious reasoning before retrieval is triggered, contrasting with meth- ods like DRAGIN that may prematurely commit to incomplete answers. In both examples shown in Figure 9, GRIP is presented with questions for which no retrieved evidence is initially available. Rather than pro- ducing a low-confidence final answer, GRIP emits an[INTERMEDIARY]token followed by an inter- nally composed partial response grounded in its parametric knowledge. For instance, when asked “what country was Slovakia?” or “who is Nebula on Guardians of the Galaxy”, GRIP provides fac- tual context (e.g., “Slovakia is a country in Central Europe” or “Nebula is a character in the Guardians franchise...”), before emitting[RETRIEVE]along with a reformulated sub-question. This pattern re- flects GRIP’s ability to recognize uncertainty, en- abling it to avoid hallucination while still making forward progress in reasoning. Such fallback be- havior illustrates a core strength of GRIP: its ca- pacity to delay final commitment until sufficient evidence is gathered. Figure 9: Example of GRIP falling back to internal knowledge and deferring retrieval when no context is available. Case 2: GRIP Refines Queries Through Iterative Retrieval Planning We present a multi-step example in which GRIP is- sues several[RETRIEVE]signals, each followed by a newly generated query. Initially, the retrieved con- tent is insufficient to support a final answer. How- ever, GRIP leverages its intermediate outputs to progressively formulate more focused and precise new queries. This behavior illustrates the model’s capacity for retrieval planning: dynamically ad- justing its information-seeking behavior based on context evolution and accumulated evidence. In this case (Figure 10), we observe that the in- termediate answer produced in the first and second rounds is nearly identical, indicating that the ini- tial retrieval step failed to surface useful content for answering the original question. Despite this, GRIP does not prematurely conclude. Instead, it formulates a more specific and context-aware query in the second round: “What is the name of the artist performing in the halftime show for the upcoming Super Bowl?” This refined question enables the retriever to access a higher-quality passage contain- ing the correct entity. In the third round, GRIP suc- cessfully emits[ANSWER]and[SOLVED], demon- strating that effective retrieval often hinges not on initial coverage but on the model’s ability to iter- atively reformulate sub-questions until the right evidence is found. This case showcases GRIP’s re- trieval planning mechanism in action, where small linguistic changes in the generated query lead to significant downstream improvements in answer- ability. Figure 10: Example of GRIP refining its retrieval query over multiple rounds to obtain a better answer. Case 3: GRIP’s Summarization Behavior by Answering with [ANSWER] In this case, we investigate how GRIP responds when explicitly instructed to finalize an answer via the[ANSWER]control token. Despite not having retrieved all necessary evidence, the model inte- grates its current reasoning trajectory and partial observations to generate a coherent summary re- sponse. This highlights GRIP’s ability to adapt to control signals and perform best-effort answering under constrained or premature termination, show- casing a key advantage of token-level behavioral regulation. Figure 11 illustrates a chained reasoning trajec- tory where GRIP first generates an incorrect in- termediate answer based on parametric memory (1880), then revises it after retrieving a more ac- curate year (1871) from external content. Rather than concluding immediately, the model goes one step further: it reformulates the question to explore the historical significance of the event, triggering another retrieval round. Finally, GRIP compresses the relevant passage into a concise factual response (1871 A.D.) using the[ANSWER]and[SOLVED] tokens. This case demonstrates GRIP’s ability to integrate evidence across turns, refine its beliefs based on new context, and summarize multi-step reasoning into a final answer—without hallucinat- ing or skipping steps. It also showcases how re- trieval can reshape the reasoning trajectory itself, not just inform it. Figure 11: Example of GRIP refining, reformulating, and compressing retrieved knowledge into a final answer using [ANSWER]. I Failure Analysis This section presents selected failure cases of GRIP to better understand its limitations in retrieval- augmented generation. While GRIP exhibits strong performance across benchmarks, we observe occa- sional failure patterns that reveal important areas for improvement. Specifically, we analyze repre- sentative cases involving incorrect final answers and redundant retrieval behaviors. Case 1: Incorrect Final Answer In some cases, GRIP successfully retrieves relevant evidence but fails to generate the correct final an- swer. This often occurs when the retrieved passages contain implicit or multi-faceted information that requires deeper logical inference. These failures highlight limitations in GRIP’s reasoning capability under answer ambiguity or insufficient summariza- tion, even when the retrieval component functions as intended. Case 2: Redundant Retrieval Steps GRIP’s token-level control enables flexible re- trieval depth, but in certain examples, the model triggers unnecessary retrievals that do not improve or even dilute answer quality. Such cases typically occur when the model fails to recognize that suffi- cient information is already available, resulting in redundant sub-queries and noisy context accumu- lation. These examples illustrate the challenges of balancing retrieval sensitivity and reasoning suffi- ciency, especially in high-recall settings. J Latent State Visualization of Control Decisions Figure 12: PCA visualization of the model’s hidden states before emitting[ANSWER](blue) and[RETRIEVE] (orange) on the NQ dataset. Each point corresponds to the hidden representation of the last token in the input prompt. The distinct clusters indicate that the model encodes separable internal states for different control decisions. To better understand the model’s internal de- cision dynamics, we visualize the hidden repre- sentations associated with the two control tokens: [ANSWER]and[RETRIEVE]. For each sample, we MethodBM25 DPR Hybrid Single-RAG (training-free)30.831.933.7 RobustRAG (training-based)33.232.133.8 GRIP41.039.241.0 Table 13: Retriever ablation on the main evaluation suite. We replace the default BM25 retriever with a dense DPR retriever and a BM25+DPR hybrid retriever (denoted Hybrid), while keeping the backbone (LLaMA3-8B), corpus (Wikipedia), and top-kpassages (k=3) fixed. Single-RAG and RobustRAG are representative training- free and training-based baselines from the main table, respectively. extract the hidden state of the last token in the in- put prompt, immediately before the model emits a control token. These hidden states represent the model’s contextual understanding just prior to trig- gering a retrieval or producing an answer. We then apply principal component analysis (PCA) to re- duce the hidden states to two dimensions and plot the results in Figure 12. The visualization reveals a clear separation be- tween the two control behaviors in the latent space. Samples that lead to[RETRIEVE]tend to cluster in a distinct region from those that lead to [ANSWER]. This suggests that the model has learned to encode different latent states for retrieval versus answer emission, reinforcing the effectiveness of control- token-based behavioral modulation. K Prompts Used in Structured Data Generation As part of our data generation pipeline, we employ GPT-based prompting to simulate how a model might revise its internal reasoning and generate improved follow-up queries. Given the original question, a partial (intermediary) answer, and a set of retrieved documents, the prompt instructs the model to refine the known fact and formulate a more targeted query that increases the likelihood of successful retrieval. This process is used to con- struct structured training data that supports GRIP’s ability to perform iterative retrieval planning in a self-supervised fashion. L Retriever Ablation To assess how sensitive our conclusions are to the choice of retriever, we evaluate GRIP and two rep- resentative baselines under three retrieval backends: sparse BM25, dense DPR, and a BM25+DPR hy- System: Given: • question: the original user question • retrieval: 3 context list of retrieved documents • reference: the model’s first attempt at a partial answer Task: 1. Re-summarize the intermediary fact that is most useful for answering the question, by combining the question, the retrieved documents, and initial_intermediary. – If initial_intermediary was incorrect or incomplete, discard or correct it. – Verify whether any retrieved document contains incorrect or misleading information based on your own knowledge, and correct the intermediary fact accordingly. – If it was correct, you may refine it. 2. Based on that finalized intermediary fact, plus the question and retrieved documents, generate a new search query that is highly likely to return evidence needed for the full answer. Output: Produce exactly one line in the following format, with no extra text: [Intermediary] <your refined known fact> [RETRIEVE] <your new query> --- Question: question Retrieved docs: docs_str Initial_intermediary: initial_intermediary Your output: Figure 13: Prompt used for data generation with GPT- 4o-mini. brid retriever. We select Single-RAG as a represen- tative training-free method and RobustRAG as a representative training-based method. For each re- triever, we keep all other factors unchanged, includ- ing the LLaMA3-8B backbone, the Wikipedia cor- pus, and the number of retrieved passages (k=3). Table 13 shows that hybrid retrieval consistently improves performance for both Single-RAG and RobustRAG, indicating that combining sparse and dense signals can provide higher-quality evidence than either retriever alone. For GRIP, DPR alone is slightly worse than BM25, while the hybrid re- triever matches the best result. Overall, GRIP re- mains substantially stronger than both baseline fam- ilies across all retriever choices, suggesting that its gains stem from the proposed token-level planning and training recipe rather than a particular retriever implementation. In the main paper we adopt BM25 as the default retriever for efficiency and controlled comparison, and report these additional retriever variants here for completeness. M Pseudocode of GRIP Decoding with Self-Triggered Information Planning The following pseudocode outlines the decoding procedure of GRIP, which integrates retrieval con- trol into autoregressive generation through self- triggered information planning. It illustrates how the model emits control tokens to decide when to re- trieve, generate sub-questions, and finalize answers during the decoding process. Algorithm 1 GRIP Decoding with Self-Triggered Information Planning Input: queryx, modelπ θ , retrieverR, max steps B Output: final answer y 1: Initialize y ←∅, m←∅, b← 0 2: while b < B do 3:Sample t∼ π θ (x,m,y) 4:Append t to y 5:if t = [ANSWER] then 6: Generate final answer until [SOLVED] 7:return y 8:else if t = [INTERMEDIARY] then 9: Generate partial answera, then gener- ate next token t ′ 10:if t ′ = [RETRIEVE] then 11:Generate new query q 12:Retrieve c←R(q) 13:m← m∪c; b← b + 1 14:else 15:Append fallback message 16:end if 17:else if t = [SOLVED] then 18:return y 19:end if 20: end while 21: return y N Retrieval-Count Distributions under Budget Control We vary the inference-time maximum retrieval bud- getB ∈3, 5, 7, 10and record the realized num- ber of retrieval callsrper example. Figure 14 re- ports the empirical distribution ofron each bench- mark. The dashed vertical line highlightsr=3, which matches the maximum retrieval depth used in our constructed training trajectories. Across all datasets, GRIP does not saturate the budget: even whenBis large, most probability mass remains on smallr(typicallyr ≤ 2), indicating that retrieval is triggered selectively when the decoding trajec- tory suggests an information gap, instead of being executed as a fixed schedule. Importantly, the figure also demonstrates extrap- olative generalization beyond the training depth. Our training trajectories are constructed with at most three retrieval steps, yet increasingBat test time yields non-trivial mass atr>3. AtB=10, the fraction of examples withr>3reaches 17.0% on WebQ, 11.7% on NQ, 7.8% on PopQA, 7.3% on HotpotQA, and 3.4% on TriviaQA. This shows that GRIP learns a transferable decision principle of when additional evidence is needed, rather than overfitting to a fixed maximum-step pattern seen during training. Finally, the extent of extrapolation is task- dependent.Datasets that are more evidence- uncertain exhibit heavier tails beyondr=3(e.g., WebQ and NQ), whereas factoid-style benchmarks concentrate more mass atr=1orr=2and rarely exceed the training depth (e.g., TriviaQA). Over- all, budget control provides a practical mechanism to trade retrieval cost for accuracy while preserv- ing GRIP’s adaptive behavior beyond the retrieval depth seen in training. O DAPO Configuration for the RL Stage We adopt the official DAPO recipe from the VERL codebase and keep its core algorithmic configura- tion unchanged; our modifications are limited to the data configuration and our task-specific reward definitions described in §2.4. 2 In particular, we enable CLIP-HIGHER, token-level policy gradient loss, dynamic sampling, and soft overlong reward shaping, and disable KL regularization (both KL- in-reward and KL loss). Soft overlong punishment is enabled and activates when the response length exceeds overlong_buffer.len. Dynamic sampling is en- abled as in the default recipe to encourage advan- tage diversity during RL training. P Results with Qwen2.5-7B To verify that our findings are not specific to a single backbone, we additionally evaluate GRIP and the reproduced baselines using Qwen2.5-7B under the same evaluation protocol. In particular, we keep the external corpus (Wikipedia), retriever (BM25), and top-kpassages (k=3) unchanged, and 2 https://github.com/volcengine/verl/blob/main/ recipe/dapo/run_dapo_qwen2.5_32b.sh 012345678910 Realized Retrieval Count (r) 0.0 0.1 0.2 0.3 0.4 Proportion WebQ 012345678910 Realized Retrieval Count (r) 0.0 0.1 0.2 0.3 0.4 0.5 Proportion PopQA 012345678910 Realized Retrieval Count (r) 0.0 0.2 0.4 0.6 Proportion NQ 012345678910 Realized Retrieval Count (r) 0.0 0.2 0.4 0.6 Proportion HotpotQA 012345678910 Realized Retrieval Count (r) 0.0 0.2 0.4 0.6 0.8 Proportion TriviaQA Max retrieval budget B B=3B=5B=7B=10 Figure 14: Realized retrieval-count distributions of GRIP under different maximum retrieval budgetsB. Each subplot corresponds to one benchmark; curves show the fraction of examples that triggerrretrieval calls at inference time. The dashed vertical line marksr=3, the maximum retrieval depth observed in our training trajectories. As Bincreases, GRIP places more probability mass beyondr>3, indicating extrapolation beyond the training depth, while the distribution remains concentrated at smallr, suggesting selective and task-aware retrieval rather than mechanically consuming the available budget. ComponentStatusKey settings CLIP-HIGHERonclip_ratio_low=0.2, clip_ratio_high=0.28, clip_ratio_c=10.0 Token-level PG lossonloss_agg_mode=token-mean Dynamic samplingonenabled (default DAPO setting) Soft overlong shapingonoverlong_buffer.len=8192, penalty_factor=1.0 KL in rewardoffuse_kl_in_reward=False, kl_coef=0.0 KL lossoffuse_kl_loss=False, kl_loss_coef=0.0 Table 14: Key DAPO components used in our RL stage. All unspecified hyperparameters follow the referenced DAPO recipe. apply the same decoding and scoring procedure as in the main experiments. The results on Qwen2.5-7B exhibit the same trend as those on LLaMA3-8B: GRIP consistently outperforms both training-free and training-based RAG baselines across datasets, while maintaining strong retrieval efficiency through token-level plan- ning. The relative ranking among baseline methods is also stable, indicating that the gains of GRIP stem from the proposed retrieval-as-generation for- mulation and scenario-typed supervision rather than backbone-specific effects. We further observe that the RL stage provides an additional but modest improvement on Qwen2.5-7B, consistent with our main findings that RL primarily refines retrieval behavior and stabilizes termination decisions. HotpotQAPopQANQWebQTriviaQA Avg. Score EM ROUGEF1EM ROUGEF1EM ROUGEF1EM ROUGEF1EM ROUGEF1 Qwen-3-4B-Instruct Instruct8.013.616.55.38.010.36.211.313.77.716.320.0 22.124.130.114.2 Single-RAG18.424.628.8 18.321.725.5 12.918.921.98.716.720.2 36.038.046.023.8 Robust-RAG 27.330.836.7 27.328.233.2 20.923.828.1 19.824.528.7 47.745.255.131.8 R1-Searcher 18.221.225.4 12.514.418.0 11.816.820.1 12.921.625.4 36.735.143.522.3 GRIP31.939.746.2 27.830.936.6 21.128.633.2 16.827.532.3 49.948.259.135.3 Qwen2.5-7B-Instruct Instruct18.321.526.0 11.513.016.3 11.816.720.0 14.522.626.4 37.136.044.422.4 Single-RAG30.634.940.9 23.626.631.1 18.924.628.3 15.322.626.3 45.846.155.731.4 Robust-RAG 26.130.835.5 27.528.633.5 20.523.827.8 21.125.830.0 48.746.756.732.2 R1-Searcher 20.224.328.8 11.613.817.0 14.319.223.0 19.326.730.7 40.139.448.325.1 GRIP32.037.644.6 28.730.936.3 21.227.732.5 19.328.333.2 51.949.660.535.6 Table 15: Main results on five QA benchmarks under two backbones. We report EM, ROUGE, F1, and Avg. Score.

Instant research discovery