Pith. sign in

REVIEW 3 major objections 5 minor 53 references

BOUND: Brief-Guided Corrective Preference Distillation at Search-Control Boundaries

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Persistent search drift in LLM-based search agents can be corrected by distilling teacher-built preferences at decision-time search-control boundaries, using a brief that pins the original target and constraints.

desk verdict BOUND is a well-controlled, credible recipe for correcting persistent search drift in LLM agents; the main risks are the unverified no-observation teacher judgments and single-run headline numbers. read the letter →

arxiv 2608.08768 v1 pith:IIUCKURM submitted 2026-08-09 cs.IR

Qingying Niu , Ruiyang Ren , Wayne Xin Zhao , Yaliang Li This is my paper · ORCID
classification cs.IR
keywords persistentsearchdriftcontrolpreferencedistillationdirectoptimizationLLMagentsdeepmulti-hopquestionansweringretrieval-augmentedgeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

BOUND claims that the persistent search drift seen in LLM-based search agents — wrong-anchor drift, dropped constraints, and local-topic drift — can be suppressed by distilling teacher-built preferences at decision-time "search-control boundaries." The core assertion is that a teacher can decide, from only the student-visible state plus a structured "search-state brief" that fixes the original target and constraints, whether the student's next move is a correctable local error, and then contrast that move with a student-specific correction (or, on successful rollouts, contrast an evidence-supported answer with an unnecessary further retrieval). These state-matched pairs are distilled into the student with direct preference optimization (DPO), while the brief and teacher stay in training. If correct, the recipe is a training-time fix that transfers to held-out deep-search tasks, with the reported gains being 5.6 EM points on Bamboogle and 4.8 accuracy points on BrowseComp-Plus over Trajectory SFT.

What carries the argument

The search-state brief: a five-field, teacher-only summary generated from the student-visible state (question, action history, evidence context) that pins the original search target and key constraints while recording confirmed evidence, missing information, and drift status. It does the work of giving the teacher a stable task-level reference so that assessment and correction are anchored to what should stay fixed, rather than to a potentially drifted context. The second piece is the state-matched preference pair — chosen and rejected continuations sharing the same state — which isolates the local search-control decision; DPO then pushes the student toward the chosen side.

What would settle it

Take a held-out set of a few hundred student decision-time states, let the student execute each continuation, and record whether the next observation actually pushes the search away from the target and constraints. Compare the teacher's error flags (made without that observation) against this ground truth. The central claim would be falsified if the teacher's precision on drift-producing states is near chance, or if retraining the student on the same pairs but with the 'correction' replaced by a randomly chosen alternative continuation of similar quality retains the benchmark gains.

Watch

Extended reading notes

Core claim

At the paper's center is a claim about what makes search drift persist: a continuation can be locally plausible — a reasonable query that retrieves relevant-looking evidence — while reinforcing a wrong anchor, a dropped constraint, or a displaced topic, so neither trajectory imitation nor outcome reward cleanly supervises the misstep. BOUND's discovery is that a teacher-side "search-state brief" (original search target, key constraints, confirmed evidence, missing information, drift status) lets the teacher separate locally plausible from task-aligned continuations using only the decision-time state, and that pairing the student's actual continuation with a brief-guided correction — or with an unnecessary retrieval step when the answer is already supported — yields preference pairs that survive validation and improve the student after DPO training. The paper reports that this transfer holds under shared retrieval and inference settings: best performance on five of six datasets, and consistent gains over a trajectory-SFT control trained with the same teacher, questions, and search-control interface.

Load-bearing premise

The load-bearing premise is that the teacher can tell, from just the student-visible state, the constructed brief, and the student's proposed next move — without seeing what that move retrieves — whether the move is a correctable local error rather than a reasonable step; if that judgment is wrong, the DPO pairs teach the student bad preferences and the reported gains could come from noise or from the teacher's own parametric knowledge.

Editorial extensions

If this is right

  • If BOUND is right, a training-time brief is enough: the student learns to avoid drift with no inference-time privileged information.
  • The corrective contrast targets the student's own misstep rather than a generic teacher move, so the signal is specific to errors that actually occur on the student's rollout distribution.
  • Termination preferences on successful rollouts teach the student to stop when evidence suffices, which should reduce oversearching and unnecessary retrieval cost.
  • The rerouting ablation implies that explicit re-anchoring — dropping passages tied to the abandoned search direction — is what makes the correction effective, not just the rewritten query.
  • The gains transfer across benchmarks (Bamboogle, BrowseComp-Plus, FRAMES, GAIA) and across two teachers, suggesting the mechanism generalizes beyond the training distribution.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • One implication the paper leaves implicit: the density of corrective pairs on a dataset should predict how much BOUND helps; measuring performance against the fraction of states flagged as drifted would test whether the preference signal, rather than the teacher's parametric knowledge, drives the gains.
  • The observation-exclusion design suggests teacher judgments should degrade as drift deepens; a controlled study varying drift magnitude would clarify the boundary of the brief's usefulness.
  • BOUND's local corrective preferences and outcome-level RL (e.g., reward on final answer) plausibly complement each other: local DPO to stop drift, outcome RL to allocate global credit, with the rerouting transition as a shared action space.
  • The 'unnecessary retrieval' contrast could be extended to rollouts that end in failure but contain a late evidence-supported answer, possibly teaching earlier stopping when the student is about to go astray.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes BOUND, a training-time preference-distillation method for LLM-based deep search agents. BOUND collects rollouts from a student policy, constructs a teacher-side 'search-state brief' for each decision-time state, and uses a teacher model to assess whether the student's continuation contains a local search-control error. Depending on the rollout outcome, it then constructs state-matched preference pairs: a corrective pair contrasting a student-specific correction with the original continuation in unsuccessful rollouts, or a termination pair contrasting an evidence-supported answer with an unnecessary retrieval in successful rollouts. These pairs are distilled into the student via DPO, with the brief and teacher computation used only at training time. The method is evaluated on HotpotQA, MuSiQue, 2WikiMultiHopQA, Bamboogle, FRAMES, GAIA, and BrowseComp-Plus, and the paper reports that BOUND outperforms Trajectory SFT by 5.6 EM on Bamboogle and 4.8 accuracy on BrowseComp-Plus under matched settings.

Significance. If the reported results hold, BOUND makes a useful contribution: it addresses a real failure mode in multi-step search (persistent drift) with an on-policy, state-matched preference construction procedure, rather than relying on full teacher trajectories or coarse outcome rewards. The paper has clear strengths: the controlled comparison with Trajectory SFT in Table 4 shares teacher, student, retriever, and evaluation protocol; the ablations in Table 3 and Figure 3 isolate specific components; the decontamination audit in Appendix C is careful; and the authors provide code. The central risk is not circularity or leakage but measurement reliability and the validity of the teacher's no-observation local-error judgment. The paper's main claims rest on single-run benchmark numbers with no variance or significance testing, and the validation of the teacher's load-bearing judgment is a small, author-conducted audit. These issues are fixable and do not invalidate the approach, but they currently limit the confidence in the reported gains.

major comments (3)
  1. [Section 4.1, Table 1] The central claim that BOUND 'achieves the best performance on five of the six datasets' is based on single-run numbers with no error bars, confidence intervals, or significance tests. Table 1 does not report how many seeds or rollouts were used, and no p-values are given for any comparison in that table. Because many reported differences between BOUND and the next-best baseline are only a few points (e.g., HotpotQA EM 42.4 versus 41.1 for R1-Searcher, MuSiQue F1 26.3 versus 24.7), the claim of consistent superiority is not statistically supported. I request repeated evaluation with multiple seeds (or bootstrap intervals) and significance tests for the main comparisons, or a clear statement of which differences are robust.
  2. [Section 3.3.1, Algorithm 1 lines 3-4; Appendix B Table 8] The load-bearing premise is that the teacher can reliably identify a correctable local search-control error from only the student-visible state, the generated brief, and the original continuation, without seeing the observation that the continuation produces. The manuscript states this explicitly in Section 3.3.1: 'The observation produced by the continuation is excluded, so the assessment relies only on information available at decision time.' The only direct check is a manual audit of 50 states by the authors, reported in Appendix B, Table 8, with 94 percent drift-status accuracy and 11 of 13 drift cases detected. This sample is small, self-annotated, and does not execute the corrections to verify that they retrieve the missing evidence under the actual rerouting transition. If the teacher mislabels locally appropriate continuations as erroneous, or if the preferred correction does not actually retrieve better evidence under the Reroute execution in Appendix D.4, the DPO training signal would be wrong. I recommend additional validation: execute a sample of teacher corrections in the environment and compare the resulting evidence and outcomes against the original continuations, and independently annotate a larger sample of teacher labels to measure inter-annotator agreement.
  3. [Section 4.4, Table 4] The controlled Trajectory SFT comparison is the strongest evidence for the method, but the statistical claim is reported only for one metric: 'The BrowseComp-Plus accuracy gain is statistically significant (p<0.001).' No significance test is reported for the Bamboogle EM/F1 gains (5.6 and 6.5 points, respectively) or for the BrowseComp-Plus recall gain. Since the abstract and Section 4.4 highlight these numbers, please clarify whether the other gains are also statistically significant, or report them as non-significant trends. At minimum, give the test procedure (number of samples, test type) used for the p-value.
minor comments (5)
  1. [Table 1 and Section 4.1] The table omits standard errors or variance, which makes it hard to assess the stability of all comparisons; adding a note about the number of evaluation runs would help readers interpret the percentages.
  2. [Figure 3] The bar chart reports four strategies without error bars or significance tests; adding confidence intervals or at least a statement about single-run evaluation would improve interpretability.
  3. [Appendix B] The manual audit is described as being done by the authors; since the audit is used to support the teacher-assessment quality, it would be better to have independent annotation or a second rater with agreement statistics, even on a small sample.
  4. [Section 3.3.2] The definition of 'search-control boundary' is distributed across Section 3.1, Section 3.3.2, and Appendix D.1; a concise formal definition placed near its first use would reduce ambiguity.
  5. [Abstract] The abstract says '12 of 14 metrics' and Section 4.2 says 'best performance on five of the six datasets'; consider aligning the phrasing and noting the 2Wiki exception explicitly in the abstract for full transparency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: BOUND's preference-construction recipe is tested on held-out benchmarks, and the paper's claims do not reduce to fitted inputs or a self-citation chain.

full rationale

BOUND is an empirical preference-distillation paper rather than a formal derivation, so there is no equation-level reduction to inspect. The teacher assesses each student continuation from the decision-time state, the constructed brief, and the continuation while explicitly excluding the resulting observation, later trajectory information, rollout outcome, and gold answers (Section 3.3.1, Appendix D.1). Preference pairs are then built from that assessment and the outcome, but the outcome only selects the contrast form and does not define the error label. Training questions come from public training splits plus synthetic construction (Appendix C), and evaluation uses held-out HotpotQA, MuSiQue, 2WikiMultiHopQA, Bamboogle, FRAMES, GAIA, and BrowseComp-Plus, including benchmarks whose training data the method does not use. The paper cites same-group work such as SimpleDeepSearcher, R1-Searcher, and FORT-Searcher, but only as baselines, experimental settings, or related work; none of these citations supplies the central premise that brief-guided corrective preferences reduce drift, which is instead supported by ablations and external benchmark comparisons. The small 50-state manual audit in Appendix B is a limitation of validation strength, not a circularity, because the audited labels are not the same quantities as the reported benchmark scores. No fitted parameter is renamed as a prediction, and no result is derived from a self-citation chain. Therefore no specific circular step can be exhibited, and the score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

BOUND is an empirical ML method, so the ledger records hand-set hyperparameters, domain assumptions about teacher reliability and stopping, and the new concepts the method introduces. No scientific constants or fitted physical parameters are involved.

free parameters (5)
  • DPO preference strength beta = 0.1
    Chosen hyperparameter for DPO training; no sensitivity analysis reported, so the reported gains depend on this value.
  • Learning rate = 1e-6
    Chosen for one-epoch DPO training; no sweep reported.
  • Reroute passage filtering threshold = 0.2
    E5-base-v2 cosine similarity cutoff for removing passages from the immediately preceding retrieval when executing Reroute (Appendix D.4); hand-set without ablation across thresholds.
  • Retrieval top-k = 5
    All methods retrieve top-5 passages; shared with baselines, so it is not a differentiator, but it constrains the results.
  • Maximum search steps = 10
    Iterative-search methods limited to 10 steps per question on multi-hop/FRAMES/GAIA; a length cap that affects which states can be corrected.
assumptions (5)
  • domain assumption Teacher-side decision-time assessment can reliably identify local search-control errors from only the state, brief, and continuation, without future observations.
    Used in AssessAndCorrect (Algorithm 1, Section 3.3.1); if the teacher's local error judgment is unreliable, preference labels are wrong. Only validated on 50 manually inspected states (Appendix B).
  • domain assumption The earliest evidence-supported answer in a successful rollout is preferable to an unnecessary retrieval continuation.
    Termination contrast construction (Algorithm 1 lines 11-17, Section 3.3.2) assumes the teacher's evidence-sufficiency judgment is correct at the stopping point.
  • standard math DPO with the initial policy as reference correctly transfers the constructed preferences to the student under the shared-state formulation.
    Relies on the DPO derivation from Rafailov et al. [26]; not re-proved here.
  • domain assumption The shared retrieval environment (BM25 plus E5-base-v2, reciprocal rank fusion) and the March 1 2022 English Wikipedia dump provide a valid and fair setting for comparing search agents.
    Underlies all baseline reruns and BOUND evaluations (Section 4.1); if retrieval differs from publication settings, comparisons could be unfair.
  • domain assumption The 983 sampled and synthetic training questions are representative enough for the method to transfer to test benchmarks.
    The method's gains transfer to Bamboogle, FRAMES, GAIA, and BrowseComp-Plus, but training questions come only from HotpotQA/MuSiQue and synthetic graphs (Appendix C).
invented entities (2)
  • Search-state brief b_t independent evidence
    purpose: Teacher-side privileged state representation with five fields (original target, key constraints, confirmed evidence, missing information, drift status); anchors assessment to task and tracks drift.
    It is an input representation, not a learned entity; its contribution is directly ablated in Table 6 and manually checked in Appendix B.
  • Search-control boundary independent evidence
    purpose: Conceptual label for states where alternative continuations differ in local search-control decisions; operationalized as a validated state-matched preference pair.
    Introduced as the unit of supervision; each valid pair is claimed to operationalize one boundary, and downstream benchmark performance provides a falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BOUND: Brief-Guided Corrective Preference Distillation at Search-Control Boundaries." pith.science (2026). https://pith.science/paper/IIUCKURM

@misc{pith2026260808768,
  author       = {Pith},
  title        = {Pith review of: BOUND: Brief-Guided Corrective Preference Distillation at Search-Control Boundaries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IIUCKURM}},
  note         = {Machine review of arXiv:2608.08768}
}
read the original abstract

Large language model (LLM)-based deep search agents solve tasks through iterative retrieval and reasoning, but locally relevant evidence can cause persistent wrong-anchor drift, constraint drift, or local-topic drift. Existing methods supervise trajectories, outcomes, or steps, but rarely distinguish task-aligned continuations from locally plausible ones that reinforce drift. We propose BOUND, a brief-guided corrective preference distillation framework for persistent search drift. For each student-induced decision-time state, BOUND constructs a teacher-side search-state brief that preserves the original search target and key constraints while summarizing confirmed evidence, missing information, and drift status. Guided by the brief, the teacher determines whether the student's continuation contains a correctable local search-control error likely to affect subsequent decisions. Together with the rollout outcome, this assessment determines whether to construct a corrective contrast between a student-specific correction and the original continuation, or a termination contrast between a supported answer and an unnecessary retrieval continuation. Each validated state-matched preference pair operationalizes a search-control boundary. Direct preference optimization (DPO) distills these preferences into the student, while the brief and teacher-side computation remain confined to training. We evaluate BOUND on four multi-hop QA benchmarks and three deep-search benchmarks. Across the six benchmarks for which we reran baselines, BOUND leads on five datasets and 12 of 14 metrics. Under the same search-control interface and matched settings, BOUND outperforms Trajectory SFT by 5.6 EM points on Bamboogle and 4.8 accuracy points on BrowseComp-Plus. Code is available at https://github.com/RUCAIBox/BOUND.

Figures

Figures reproduced from arXiv: 2608.08768 by the authors.

Figure 1
Figure 1. Three manifestations of persistent search drift and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of BOUND. A teacher-side brief and rollout outcomes yield validated state-matched preference pairs that [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Effect of preference-pair construction strategies. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 16 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. GPT-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862(2022)

    arXiv 2022

  3. [3]

    Zijian Chen, Xueguang Ma, Shengyao Zhuang, Ping Nie, Kai Zou, Andrew Liu, Joshua Green, Kshama Patel, Ruoxi Meng, Mingyi Su, et al. 2025. BrowseComp- Plus: A more fair and transparent evaluation benchmark of deep-research agent. arXiv preprint arXiv:2508.06600(2025)

    arXiv 2025

  4. [4]

    Gordon V Cormack, Charles LA Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval. 758–759

    2009

  5. [5]

    Lu Dai, Liang Sun, Fanpu Cao, Ziyang Rao, Cehao Yang, Hao Liu, and Hui Xiong

    work page

  6. [6]

    DeepSeek-AI. 2026. DeepSeek-V4: Towards Highly Efficient Million-Token Con- text Intelligence

    work page 2026

  7. [7]

    Jia Deng, Yimeng Chen, Xiaoqing Xiang, Ziyang Zeng, Shuo Tang, Wayne Xin Zhao, Feng Chang, Chuan Hao, Yuan Wei, Ran Tao, et al. 2026. FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents. arXiv preprint arXiv:2606.12087(2026)

    work page Pith review arXiv 2026

  8. [8]

    Namgyu Ho, Laura Schmid, and Se-Young Yun. 2023. Large Language Models Are Reasoning Teachers. InProceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers). 14852–14882

    work page 2023

Show all 53 references
  1. [9]

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi-hop qa dataset for comprehensive evaluation of reason- ing steps. InProceedings of the 28th International Conference on Computational Linguistics. 6609–6625

    2020

  2. [10]

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. 2023. Distilling Step- by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes. InFindings of the As...

    2023

  3. [11]

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-R1: Training LLMs to rea- son and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516(2025)

    2025 arXiv

  4. [12]

    Rikuto Kotoge, Mai Nishimura, and Jiaxin Ma. 2026. Can Compact Language Models Search Like Agents? Distillation-Guided Policy Optimization for Preserv- ing Agentic RAG Capabilities. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volu...

    2026

  5. [13]

    Satyapriya Krishna, Kalpesh Krishna, Anhad Mohananey, Steven Schwarcz, Adam Stambler, Shyam Upadhyay, and Manaal Faruqui. 2025. Fact, fetch, and reason: A unified evaluation of retrieval-augmented generation. InProceedings of the 2025 Conference of the Nations of the Americas ...

    2025

  6. [14]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. InAdvances in N...

    2020

  7. [15]

    Kuan Li, Zhongwang Zhang, Huifeng Yin, Liwen Zhang, Litu Ou, Jialong Wu, Wenbiao Yin, Baixuan Li, Zhengwei Tao, Xinyu Wang, Weizhou Shen, Junkai Zhang, Dingchu Zhang, Xixi Wu, Yong Jiang, Ming Yan, Pengjun Xie, Fei Huang, and Jingren Zhou. 2025. WebSailor: Navigating Super-hum...

    2025 arXiv

  8. [16]

    Mukai Li, Qingcheng Zeng, Tianqing Fang, Zhenwen Liang, Linfeng Song, Qi Liu, Haitao Mi, and Dong Yu. 2026. Verified Critical Step Optimization for LLM Agents. InFindings of the Association for Computational Linguistics: ACL 2026, Maria Liakata, Viviane P. Moreira, Jiajun Zhan...

    2026

  9. [17]

    Minghan Li and Guodong Zhou. 2026. Retrieval-Feedback-Driven Distillation and Preference Alignment for Efficient LLM-based Query Expansion.arXiv preprint arXiv:2603.13776(2026)

    2026

  10. [18]

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. 2025. Search-o1: Agentic search-enhanced large reasoning models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 5420–5438

    2025

  11. [19]

    Xiaozhe Li, Tianyi Lyu, Yang Li, Yichuan Ma, Peiji Li, Linyang Li, Qipeng Guo, Dahua Lin, and Kai Chen. 2026. What and When to Distill: Selective Hindsight Distillation for Multi-Turn Agents.arXiv preprint arXiv:2605.19447(2026)

    2026 arXiv

  12. [20]

    Yufei Ma, Zihan Liang, Ben Chen, Zhipeng Qian, Huangyu Dai, Lingtao Mao, Xuxin Zhang, Chenyi Lei, and Wenwu Ou. 2026. SD-Search: On-Policy Hindsight Self-Distillation for Search-Augmented Reasoning.arXiv preprint arXiv:2605.18299(2026)

    2026 arXiv

  13. [21]

    Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. 2024. GAIA: a benchmark for General AI Assistants. In International Conference on Learning Representations

    2024

  14. [22]

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al

  15. [23]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schul- man, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike,...

    2022

  16. [24]

    Emiliano Penaloza, Dheeraj Vattikonda, Nicolas Gontier, Alexandre Lacoste, Laurent Charlin, and Massimo Caccia. 2026. Privileged Information Distillation for Language Models.arXiv preprint arXiv:2602.04942(2026)

    2026 arXiv

  17. [25]

    Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. InFindings of the Association for Computational Linguistics: EMNLP 2023. 5687–5711

    2023

  18. [26]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. InAdvances in Neural Information Processing Systems, Vol. 36. Curran Associates, Inc., 5...

    2023 doi

  19. [27]

    Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Frame- work: BM25 and Beyond.Foundations and Trends in Information Retrieval3 (09 2009), 333–389. doi:10.1561/1500000019

    2009 doi

  20. [28]

    Idan Shenfeld, Mehul Damani, Jonas Hübotter, and Pulkit Agrawal. 2026. Self- Distillation Enables Continual Learning.arXiv preprint arXiv:2601.19897(2026)

    2026 arXiv

  21. [29]

    Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2023. Distilling reasoning capabilities into smaller language models. InFindings of the Association for Computational Linguistics: ACL 2023. 7059–7073

    2023

  22. [30]

    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 preprint arXiv:2503.05592 (2025)

    2025 arXiv

  23. [31]

    Shuang Sun, Huatong Song, Yuhao Wang, Ruiyang Ren, Jinhao Jiang, Jun- jie Zhang, Fei Bai, Jia Deng, Wayne Xin Zhao, Zheng Liu, et al . 2025. Sim- pleDeepSearcher: Deep Information Seeking via Web-Powered Reasoning Tra- jectory Synthesis.arXiv preprint arXiv:2505.16834(2025)

    2025

  24. [32]

    Qwen Team. 2026. Qwen3.5: Accelerating Productivity with Native Multimodal Agents. https://qwen.ai/blog?id=qwen3.5 Conference’17, July 2017, Washington, DC, USA Niu et al

    2026

  25. [33]

    Nandan Thakur, Zijian Chen, Xueguang Ma, and Jimmy Lin. 2026. ORBIT: Scalable and Verifiable Data Generation for Search Agents on a Tight Budget. arXiv preprint arXiv:2604.01195(2026)

    2026

  26. [35]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal

  27. [36]

    Hao Wang, Guozhi Wang, Han Xiao, Yufeng Zhou, Yue Pan, Jichao Wang, Ke Xu, Yafei Wen, Xiaohu Ruan, Xiaoxin Chen, et al. 2026. Skill-SD: Skill-conditioned self-distillation for multi-turn llm agents.arXiv preprint arXiv:2604.10674(2026)

    2026 arXiv

  28. [37]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. Text embeddings by weakly-supervised contrastive pre-training.arXiv preprint arXiv:2212.03533(2022)

    2022 arXiv

  29. [38]

    Ziliang Wang, Xuhui Zheng, Kang An, Cijun Ouyang, Jialu Cai, Yuhang Wang, and Yichao Wu. 2025. StepSearch: Igniting llms search ability via step-wise proximal policy optimization.arXiv preprint arXiv:2505.15107(2025)

    2025 arXiv

  30. [39]

    Tongyu Wen, Guanting Dong, and Zhicheng Dou. 2026. SmartSearch: Pro- cess Reward-Guided Query Refinement for Search Agents.arXiv preprint arXiv:2601.04888(2026)

    2026

  31. [40]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    2025 arXiv

  32. [41]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. InProceedings of the 2018 conference on empirical methods in natural language p...

    2018

  33. [42]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. InInternational Conference on Learning Representations (ICLR)

    2023

  34. [43]

    Woongyeng Yeo, Yumin Choi, Taekyung Ki, and Sung Ju Hwang. 2026. HINT-SD: Targeted Hindsight Self-Distillation for Long-Horizon Agents.arXiv preprint arXiv:2605.17873(2026)

    2026 arXiv

  35. [44]

    Rongzhi Zhang, Jiaming Shen, Tianqi Liu, Haorui Wang, Zhen Qin, Feng Han, Jialu Liu, Simon Baumgartner, Michael Bendersky, and Chao Zhang. 2024. PLaD: Preference-based Large Language Model Distillation with Pseudo-Preference Pairs. InFindings of the Association for Computation...

    2024 doi

  36. [45]

    Wenlin Zhang, Kuicai Dong, Junyi Li, Yingyi Zhang, Xiaopeng Li, Pengyue Jia, Yi Wen, Derong Xu, Maolin Wang, Yichao Wang, et al. 2026. To search or not to search: Aligning the decision boundary of deep search agents via causal intervention. InProceedings of the ACM Web Confere...

    2026

  37. [46]

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou

  38. [47]

    Yanfei Zhang, Xu Lin, and Chenglin Wu. 2026. StepOPSD: Step-Aware On- line Preference Distillation for Agent Reinforcement Learning.arXiv preprint arXiv:2605.27140(2026)

    2026 arXiv

  39. [48]

    Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. 2026. Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models.arXiv preprint arXiv:2601.18734(2026)

    2026 arXiv

  40. [49]

    search history

    Qiyong Zhong, Mao Zheng, Mingyang Song, Xin Lin, Jie Sun, Houcheng Jiang, Xiang Wang, and Junfeng Fang. 2026. SOD: Step-wise on-policy distillation for small language model agents.arXiv preprint arXiv:2605.07725(2026). BOUND: Brief-Guided Corrective Preference Distillation at ...

    2026 arXiv

  41. [2021]

    arXiv preprint arXiv:2112.09332(2021)

    WebGPT: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332(2021)

    2021 arXiv

  42. [2022]

    Transactions of the Association for Computational Linguistics10 (2022), 539–554

    MuSiQue: Multihop Questions via Single-hop Question Composition. Transactions of the Association for Computational Linguistics10 (2022), 539–554

    2022

  43. [2023]

    InProceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers)

    Interleaving retrieval with chain-of-thought reasoning for knowledge- intensive multi-step questions. InProceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long papers). 10014–10037

  44. [2025]

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.arXiv preprint arXiv:2506.05176(2025)

    2025 arXiv

  45. [2026]

    LLM-Oriented Information Retrieval: A Denoising-First Perspective.arXiv preprint arXiv:2605.00505(2026)

    2026 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.