Pith. sign in

REVIEW 4 major objections 7 minor 4 cited by

EvolveSearch: An Iterative Self-Evolving Search Agent

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 7-billion-parameter web-search agent can bootstrap its own skill by alternating reinforcement learning with supervised fine-tuning on its own best rollouts, reaching a 4.7% average gain over prior state of the art on seven multi-hop QA…

desk verdict A clean iterative RL+SFT loop for web-search agents that likely improves accuracy, but the 4.7% headline is measured by the same judge used for training and filtering, and alternate judges shrink the gap to 1-3 points. read the letter →

arxiv 2505.22501 v1 pith:TQCITY54 submitted 2025-05-28 cs.CL

Dingchu Zhang , Yida Zhao , Jialong Wu , Baixuan Li , Wenbiao Yin , Liwen Zhang , Yong Jiang , Yufeng Li
show 3 more authors
Kewei Tu Pengjun Xie Fei Huang
This is my paper · ORCID
classification cs.CL
keywords iterativeself-evolutionwebsearchagentsreinforcementlearningrejectionsamplingfine-tuningmulti-hopquestionansweringGRPOhybridrewardtool-uselanguagemodels
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

EvolveSearch claims that a language-model agent can teach itself web search without human-written reasoning traces. It alternates reinforcement learning (RL), where the agent searches and gets rewarded for well-formatted, correct answers, with rejection-sampling fine-tuning, where the best and most tool-heavy rollouts from RL are used as supervised training data. The new fine-tuned model becomes the starting point for the next RL round, so each cycle starts stronger than the last. On seven multi-hop question-answering benchmarks, the paper reports a 4.7% average accuracy gain over the previous best method, and its control experiment shows RL alone improves less and stalls while the alternating loop keeps climbing.

What carries the argument

The load-bearing object is the iterative data pool. The raw data is split into N parts; in each iteration the current cold-start policy is trained with a token-level variant of GRPO on one part, with reward $R = 0.5(R_f + R_a)$, where $R_f$ is 1.0 for exact format compliance and $R_a$ is 1.0 when a judge model certifies the answer correct. Rollouts are then filtered by three rules—reward at least 0.7, same-query deduplication keeping the most tool-heavy rollout, and selection of the top-k rollouts by tool-call count—and accumulated with previous iterations' data. The base model is SFT-trained on this filtered pool with tool-response tokens masked out, yielding the cold-start model for the next RL round. This stack lets RL's exploration output feed back into supervised learning, which is what sustains multi-iteration gains.

What would settle it

Have human annotators, or a set of judge models never exposed to the training loop, score the same test answers on all seven benchmarks. If the EvolveSearch advantage over RL-only and over the strongest baseline shrinks to the roughly one-to-three-point range seen in the paper's alternate-judge table, or disappears, the central claim of genuine test-accuracy improvement is refuted.

Watch

Extended reading notes

Core claim

The central discovery is that high-reward rollouts generated during RL exploration are a reusable supervised signal rather than a dead end: filtering them by reward, query diversity, and number of tool calls, then running supervised fine-tuning (SFT) on the accumulated pool, produces a better cold-start policy for the next RL phase. This loop is the reason performance improves with iteration count. After three iterations EvolveSearch reports an average of 76.2% in-domain and 53.7% out-of-domain across seven MHQA benchmarks, beating all listed baselines and improving over its own RL-only counterpart, which reaches 73.8% in-domain and 53.2% out-of-domain at iteration three. The paper claims this establishes a self-evolution mechanism for open-domain web search that needs no external reasoning annotations.

Load-bearing premise

The entire reported gain rests on the assumption that the same 72-billion-parameter judge model used to give RL rewards, filter SFT data, and grade the test answers is a valid, unbiased judge of correctness; the paper's own alternate-judge table shows the advantage shrinking to a few points, so if the agent has learned to please this judge rather than answer truthfully, the 4.7% headline overstates real improvement.

Editorial extensions

If this is right

  • If the claim holds, a practical recipe follows: collect RL rollouts, filter out low-reward and repetitive ones, keep multi-call searches, and use them as SFT data; this should keep an agent improving across more iterations than pure RL.
  • The filtering rules are each load-bearing: removing any one hurts both in-domain and out-of-domain results, so the gain is not from simply adding more SFT data.
  • The method is robust to the choice of answer reward: switching from a judge model to recall or F1 keeps the relative improvement, meaning the self-evolution loop, not the reward scale, drives the gain.
  • The reported out-of-domain gains on Musique, Bamboogle, and PopQA indicate the agent learns transferable search-and-answer behavior rather than memorized training queries.

Reading between the lines

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

  • Editorial inference: the same alternating RL-plus-filtered-SFT loop should transfer to other sparse-reward tool-use tasks, such as code execution or database queries, where reward can be checked automatically.
  • The paper's quality-over-quantity result suggests a testable refinement: instead of a fixed top-k by tool-call count, select SFT rollouts by marginal estimated value, which might sustain improvements past the three iterations shown.
  • A harder test of self-evolution would use data from rollouts the model could not already produce at the current iteration; the reported gains may partly come from consolidating behaviors already present in the RL distribution.
  • Because the judge used for rewards, data filtering, and final evaluation is the same model family, the headline gain could partly reflect reward hacking; inserting periodic human spot-checks into the reward loop would tell whether the 4.7% gain is real answer-quality improvement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes EvolveSearch, an iterative framework for training a web-search agent by alternating GRPO reinforcement learning with rejection-sampling supervised fine-tuning (RSFT). In each iteration, RL rollouts with high hybrid reward (format reward plus answer reward) are filtered by three rules (high reward, same-query deduplication, most tool calls) and used to fine-tune a fixed base model, Qwen2.5-7B-Instruct, which then becomes the cold-start for the next RL iteration. The method is evaluated on seven multi-hop QA benchmarks (NQ, TQ, HotpotQA, 2Wiki in-domain; Musique, Bamboogle, PopQA out-of-domain) using Qwen2.5-72B-Instruct as the answer judge. The authors report consistent gains over prior agents and an average improvement of 4.7% over the previous state of the art, with ablations showing each filtering rule matters, robustness to different answer-reward metrics, and qualitatively similar trends under three alternative judge models. The central claim is that repeated alternating RL and SFT on self-generated filtered rollouts yields a genuinely more accurate web-search agent without human-annotated reasoning data.

Significance. If the central claim holds, the paper makes a useful empirical contribution: it offers a simple and effective recipe for combining RL exploration with SFT exploitation in an iterative loop, addressing a known limitation of RL-only training for web-search agents (quick convergence and low data efficiency). The ablations in Section 5 support the importance of the three filtering rules, and the experiments with different answer rewards and different judge models are honest attempts to probe robustness. The method requires no external human-annotated reasoning data, which is a practically valuable property. However, the strength of the evidence as presented is limited by the coupling between the model that provides the reward, the SFT filter, and the main evaluation judge, as well as by the absence of human evaluation, confidence intervals, and documented train/eval deduplication. The paper would be significantly stronger if these gaps were addressed; as it stands, the headline gains are plausible but not yet fully convincing as a measure of true QA ability.

major comments (4)
  1. [Section 4.1, Section 3.1.1, Section 3.2 (Rule 1)] The main evaluation uses Qwen2.5-72B-Instruct as the answer judge, and this same model supplies the answer reward in RL and the high-reward filter for SFT rollouts. The policy is therefore optimized to satisfy the specific judge used to report the headline numbers. Table 3, which reports results with DeepSeek-V3, chatgpt-4o-latest, and grok-3 as judges, is reassuring in that the trend persists, but the gaps shrink considerably (roughly 1.3–1.7 points in-domain and 2.8–3.6 points out-of-domain), and all three are still LLM judges with unknown and possibly correlated false-acceptance patterns. Since the central claim is about real accuracy gain, please add a human-scored sample of the evaluation sets (or an exact-match subset) and report agreement between each LLM judge and human labels. If a full human evaluation is not feasible, at least report per-sample agreement and confidence intervals for the alternate judges so readers can gauge the sensitivity of the 4.7% figure to the judge choice.
  2. [Abstract, Table 1] The headline 'average improvement of 4.7% over the current state-of-the-art' appears to be computed against DeepResearcher (the F1-reward baseline), not against the stronger DeepResearcher* (model-based reward) baseline presented in the same table. Against DeepResearcher*, the average gain across the seven benchmarks is approximately 3.2 points, and the gain on PopQA is only 0.1 points (50.3 vs 50.2). The abstract and Section 4.4 should state which baseline is considered the previous SOTA, and the 4.7% claim should be reconciled with the comparison to the strongest baseline.
  3. [Section 4.1] In-domain evaluation draws from the same datasets used to build the 80k training mixture (NQ, TQ, HotpotQA, 2Wiki). The paper does not document any deduplication between training and test questions. Since RL generates 16 rollouts per training question and SFT preferentially selects high-reward and high-tool-use rollouts, exact or near-duplicate questions in the test set could inflate the in-domain scores. Please report an overlap check (e.g., exact-match or embedding-based deduplication) and, if any overlap exists, re-run the in-domain evaluation on a deduplicated subset to establish that the reported improvements are not partly attributable to memorization.
  4. [Section 4.4, Tables 1–4, Figure 4] All reported results are single-run point estimates with no standard deviations, confidence intervals, or significance tests. Several differences that carry the main claim are small, such as EvolveSearch-ite3 vs RLSearch-ite3 on Bamboogle (77.1 vs 77.0) and EvolveSearch-ite3 vs DeepResearcher* on PopQA (50.3 vs 50.2). Additionally, the statement that performance 'consistently improves' across iterations is contradicted by drops from iteration 2 to iteration 3 on 2Wiki (78.5 to 76.4) and PopQA (52.8 to 50.3). Please report multiple seeds or at least paired bootstrap confidence intervals for the final models, and soften claims of consistency to match the observed trajectories.
minor comments (7)
  1. [Section 4.2] The data is split into N=10 parts, but the main results report only 3 iterations. Please clarify how many of the 10 parts are used in the reported runs and how the final model is selected.
  2. [Table 3] The column headers 'A VG ID' and 'A VGOOD' appear to be formatting errors for 'Avg ID' and 'Avg OOD'; they should be corrected for readability.
  3. [Section 3.1.2, Eq. (3)] The notation in Eq. (3) is incomplete: the distribution over x is unstated, the reference policy π_ref is not defined, and the D_KL term is not specified. Please define these quantities explicitly.
  4. [Section 3.2, Rule 3] The rule says 'we select the top k rollouts with the most tool calls for SFT in each iteration,' but it is unclear whether k is a per-iteration quota, a per-query quota, or a global pool size, and how this interacts with Rule 2's deduplication. Please clarify the exact selection procedure.
  5. [Equation (5)] The loss in Eq. (5) writes y_i ∈ {τ, α, o}, but the tokens are then indexed inside the same expression; please clarify the distinction between rollout-level symbols and token-level loss positions.
  6. [References] Some references are incomplete or informal, e.g., 'Alpaca. 2023', 'Llama-3. 2024', and several arXiv entries lack version numbers or venue information; please standardize the bibliography.
  7. [Limitations] The Limitations section discusses computational cost and the restriction to a single tool, but it does not mention the judge-model coupling or the lack of train/eval deduplication. These are the most consequential limitations of the current evidence and should be explicitly acknowledged and addressed.

Circularity Check

1 steps flagged · score 4.0 of 10

Headline 4.7% gain is measured by the same Qwen2.5-72B-Instruct judge that supplies the RL answer reward and the SFT filter; Table 3 gives partial independent support.

  1. fitted input called prediction [Section 3.1.1 (Eq. 2), Section 3.2 Rule 1, Section 4.1-4.2, Table 1]
    ""Answer Reward. When the format is strictly followed, we employ a judge model to assess the correctness of the answer... A rollout receives a 1.0 reward (Ra = 1.0) if the answer is correct, otherwise Ra = 0.0." "Rule 1: High-Reward Selection (HRS). We only retain rollouts with rewards ≥ δ." "We utilize Qwen2.5-72B-Instruct as our judge model." "We utilize a judge model to evaluate the correctness of the model's response.""

    The same Qwen2.5-72B-Instruct model defines Ra in the RL reward (Eq. 2), selects SFT training data through Rule 1 (δ = 0.7), and is the evaluator for the main results (Secs. 4.1-4.2, Table 1). Hence the reported '4.7% average improvement' is an average of the very judge score the policy was trained to maximize and the filter was designed to select. A policy can inflate this number by conforming to that judge's biases without becoming more factually correct. Table 3 re-judges with DeepSeek-V3, chatgpt-4o-latest, and grok-3, where the gain narrows to roughly 1-3.6 points, showing some independent content, but no human scoring is reported. The headline number is therefore not independent of the training signal.

full rationale

EvolveSearch's alternating RL/RSFT loop is a legitimate self-training mechanism; using self-generated rollouts is not circular by itself. The circularity is in the measurement chain: the answer reward Ra, the SFT filter threshold, and the main evaluation metric are all supplied by one Qwen2.5-72B-Instruct judge. This makes the flagship improvement partly a measure of fit to that judge rather than to an independent correctness standard. The alternate-judge experiment (Table 3) provides genuine but weaker corroboration (gains shrink to about 1.3-3.6 points), so the central claim still has independent content; I therefore score this as moderate partial circularity (4), not full collapse. I found no load-bearing self-citation or imported-uniqueness issue: related-work citations from overlapping authors are not used to justify the method. Other concerns, such as single-seed runs and no released deduplication check, are statistical-validity risks rather than circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a small set of hand-chosen hyperparameters, the validity of the LLM judge as an accuracy oracle, and the assumption that high-reward, multi-tool rollouts transfer to better behavior. No new entities are postulated. The main burden is that the judge is used on both sides of the training and evaluation loop.

free parameters (5)
  • Reward threshold delta = 0.7
    Used in Rule 1 to decide which RL rollouts are kept for SFT; set by hand with no sensitivity analysis reported.
  • SFT selection budget k = 2000
    Top-k rollouts by number of tool calls are selected each iteration (Rule 3); no ablation of k is reported.
  • Hybrid reward weights = 0.5 * Rf + 0.5 * Ra
    The composition of format and answer rewards in Eq. 2 is chosen by hand.
  • Number of data splits N = 10
    Training data is divided into 10 parts, one RL phase per part; no analysis of sensitivity to N is given.
  • RL rollout count = 16 rollouts per sample
    16 rollouts per training input at temperature 1.0; no sensitivity analysis is reported.
assumptions (4)
  • domain assumption An LLM judge's binary Correct/Incorrect verdict is a valid measurement of answer correctness.
    Used both for the RL answer reward (Section 3.1.1) and for the official test evaluation (Section 4.1), with only other LLM judges as validation in Table 3.
  • ad hoc to paper RL rollouts that score high reward and contain many tool calls are high-quality training data for the next SFT phase.
    This is the core inductive bias of Rules 1 to 3 in Section 3.2; it is asserted and partially ablated but not derived.
  • domain assumption The token-level GRPO loss with KL reference (Eq. 3) improves the policy as intended in this environment.
    Taken from prior RL work (Shao et al., 2024; Yu et al., 2025), applied without a value model or further validation in this setting.
  • domain assumption Training and evaluation data from the same four benchmarks are disjoint despite both being sampled from public datasets.
    Section 4.1 reuses 'the same training and testing data as DeepResearcher' but does not document deduplication between the 80k training samples and the 2,048 evaluation examples.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EvolveSearch: An Iterative Self-Evolving Search Agent." pith.science (2026). https://pith.science/paper/TQCITY54

@misc{pith2026250522501,
  author       = {Pith},
  title        = {Pith review of: EvolveSearch: An Iterative Self-Evolving Search Agent},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQCITY54}},
  note         = {Machine review of arXiv:2505.22501}
}
read the original abstract

The rapid advancement of large language models (LLMs) has transformed the landscape of agentic information seeking capabilities through the integration of tools such as search engines and web browsers. However, current mainstream approaches for enabling LLM web search proficiency face significant challenges: supervised fine-tuning struggles with data production in open-search domains, while RL converges quickly, limiting their data utilization efficiency. To address these issues, we propose EvolveSearch, a novel iterative self-evolution framework that combines SFT and RL to enhance agentic web search capabilities without any external human-annotated reasoning data. Extensive experiments on seven multi-hop question-answering (MHQA) benchmarks demonstrate that EvolveSearch consistently improves performance across iterations, ultimately achieving an average improvement of 4.7\% over the current state-of-the-art across seven benchmarks, opening the door to self-evolution agentic capabilities in open web search domains.

Figures

Figures reproduced from arXiv: 2505.22501 by the authors.

Figure 1
Figure 1. Iterative improvements in the average perfor [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of EvolveSearch. We iteratively input raw question-answering data and a fixed base model. In the i-th iteration, the base model first performs SFT with a filter pool of training data, followed by RL with the i-th raw data. The rollout data during RL is filtered and incorporated into the training data pool for SFT in the subsequent iteration. <think> To find out who the screenwriter for The Blac… view at source ↗
Figure 3
Figure 3. The illustration of a rollout that exactly [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The impact of different data filtering rules on [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Performance of SFT model and RL model at [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Average number of tool calls in the test set [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: The distribution of tool calls for RL rollouts. [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Fetch-then-Explore: Decoupling Selection from Extraction over a Persistent Workspace for Search Agents

    cs.AI 2026-08 conditional novelty 6.0 of 10

    Fetch-then-Explore, which stores fetched pages in a per-question workspace and extracts evidence on demand with grep/read, beats visit-and-read and browsing baselines on BrowseComp across three LLM backbones.

  2. LocalSearchBench: Benchmarking Agentic Search in Real-World Local Life Services

    cs.AI 2025-12 conditional novelty 6.0 of 10

    LocalSearchBench—1.3M merchant records and 900 multi-hop local-life QA tasks across 9 Chinese cities—shows the best reasoning agent reaches only 35.6% correctness.

  3. Chain-of-Agents: End-to-End Agent Foundation Models via Multi-Agent Distillation and Agentic RL

    cs.AI 2025-08 conditional novelty 6.0 of 10

    A single LLM is trained with multi-agent distilled trajectories plus agentic RL, and the resulting Chain-of-Agents models set state-of-the-art Pass@1 scores among tool-integrated reasoning methods on GAIA, BrowseComp,...

  4. WebShaper: Agentically Data Synthesizing via Information-Seeking Formalization

    cs.CL 2025-07 conditional novelty 6.0 of 10

    WebShaper formalizes web information-seeking tasks as set-theoretic queries, synthesizes training questions by layer-wise expansion, and uses the resulting data to train open-source agents that outperform prior open-s...

Reference graph

Works this paper leans on

46 extracted references · 11 canonical work pages · cited by 4 Pith papers

  1. [1]

    Alpaca. 2023. https://crfm.stanford.edu/2023/03/13/alpaca.html Introducing alpaca: A strong and performant instruction-following language model . Accessed: 2024-06-10

    work page 2023

  2. [2]

    Reza Yazdani Aminabadi, Samyam Rajbhandari, Minjia Zhang, Ammar Ahmad Awan, Cheng Li, Du Li, Elton Zheng, Jeff Rasley, Shaden Smith, Olatunji Ruwase, and Yuxiong He. 2022. https://arxiv.org/abs/2207.00032 Deepspeed inference: Enabling efficient inference of transformer models at unprecedented scale . Preprint, arXiv:2207.00032

    arXiv 2022

  3. [3]

    Anthropic. 2025. https://www.anthropic.com/engineering/building-effective-agents Building effective agents

    work page 2025

  4. [4]

    Pan, Wen Zhang, Huajun Chen, Fan Yang, Zenan Zhou, and Weipeng Chen

    Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z. Pan, Wen Zhang, Huajun Chen, Fan Yang, Zenan Zhou, and Weipeng Chen. 2025. https://arxiv.org/abs/2503.19470 Research: Learning to reason with search for llms via reinforcement learning . Preprint, arXiv:2503.19470

    arXiv 2025

  5. [5]

    DeepSeek-AI. 2024. https://arxiv.org/abs/2412.19437 Deepseek-v3 technical report . Preprint, arXiv:2412.19437

    arXiv 2024

  6. [6]

    Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Jingyi Song, and Hao Wang. 2025. Airrag: Activating intrinsic reasoning for retrieval augmented generation via tree-based search. arXiv preprint arXiv:2501.10053

    arXiv 2025

  7. [7]

    Junxian He, Jiatao Gu, Jiajun Shen, and Marc'Aurelio Ranzato. 2019. Revisiting self-training for neural sequence generation. arXiv preprint arXiv:1909.13788

    arXiv 2019

  8. [8]

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. https://arxiv.org/abs/2011.01060 Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps . Preprint, arXiv:2011.01060

    arXiv 2020

Show all 46 references
  1. [9]

    Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander 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. arXiv preprint a...

    2023 arXiv

  2. [10]

    Jiaxin Huang, Shixiang Shane Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2022. Large language models can self-improve. arXiv preprint arXiv:2210.11610

    2022 arXiv

  3. [11]

    Yeo Wei Jie, Ranjan Satapathy, and Erik Cambria. 2024. Plausible extractive rationalization through semi-supervised entailment signal. arXiv preprint arXiv:2402.08479

    2024 arXiv

  4. [12]

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. Search-r1: Training llms to reason and leverage search engines with reinforcement learning. arXiv preprint arXiv:2503.09516

    2025 arXiv

  5. [13]

    Weld, and Luke Zettlemoyer

    Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. 2017. https://arxiv.org/abs/1705.03551 Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension . Preprint, arXiv:1705.03551

    2017 arXiv

  6. [14]

    Satyapriya Krishna, Jiaqi Ma, Dylan Slack, Asma Ghandeharioun, Sameer Singh, and Himabindu Lakkaraju. 2024. Post hoc explanations of language models can improve language models. Advances in Neural Information Processing Systems, 36

    2024

  7. [15]

    Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav...

    2019

  8. [16]

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. 2025. https://doi.org/10.48550/ARXIV.2501.05366 Search-o1: Agentic search-enhanced large reasoning models . CoRR, abs/2501.05366

    2025 doi arXiv

  9. [17]

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. 2025. Understanding r1-zero-like training: A critical perspective. arXiv preprint arXiv:2503.20783

    2025 arXiv

  10. [18]

    Llama-3. 2024. https://ai.meta.com/blog/meta-llama-3/ Meta llama 3 . Accessed: 2024-06-10

    2024

  11. [19]

    Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Hannaneh Hajishirzi, and Daniel Khashabi. 2022. When not to trust language models: Investigating effectiveness and limitations of parametric and non-parametric memories. arXiv preprint

    2022

  12. [20]

    Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. 2022. Generating training data with language models: Towards zero-shot language understanding. Advances in Neural Information Processing Systems, 35:462--477

    2022

  13. [21]

    OpenAI. 2025. https://cdn.openai.com/deep-research-system-card.pdf Deep research system card

    2025

  14. [22]

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

    2022

  15. [23]

    Melissa Z Pan, Mert Cemri, Lakshya A Agrawal, Shuyi Yang, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Kannan Ramchandran, Dan Klein, and 1 others. 2025. Why do multiagent systems fail? In ICLR 2025 Workshop on Building Trust in Language Models and Applications

    2025

  16. [24]

    Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. 2022. Measuring and narrowing the compositionality gap in language models. arXiv preprint arXiv:2210.03350

    2022 arXiv

  17. [25]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  18. [26]

    Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Ji-Rong Wen, Yang Lu, and Xu Miu. 2025. https://github.com/RUCAIBox/R1-searcher R1-searcher: Incentivizing the search capability in llms via reinforcement learning

    2025

  19. [27]

    Katherine Tian, Eric Mitchell, Huaxiu Yao, Christopher D Manning, and Chelsea Finn. 2023. Fine-tuning language models for factuality. arXiv preprint arXiv:2311.08401

    2023 arXiv

  20. [28]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. Musique: Multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10:539--554

    2022

  21. [29]

    Lewis Tunstall, Edward Beeching, Nathan Lambert, Nazneen Rajani, Kashif Rasul, Younes Belkada, Shengyi Huang, Leandro von Werra, Cl \'e mentine Fourrier, Nathan Habib, and 1 others. 2023. Zephyr: Direct distillation of lm alignment. arXiv preprint arXiv:2310.16944

    2023 arXiv

  22. [30]

    Liang Wang, Haonan Chen, Nan Yang, Xiaolong Huang, Zhicheng Dou, and Furu Wei. 2025. Chain-of-retrieval augmented generation. arXiv preprint arXiv:2501.14342

    2025

  23. [31]

    Ziting Wang, Haitao Yuan, Wei Dong, Gao Cong, and Feifei Li. 2024. Corag: A cost-constrained retrieval optimization system for retrieval-augmented generation. arXiv preprint arXiv:2411.00744

    2024

  24. [32]

    Jason Wei, Nguyen Karina, Hyung Won Chung, Yunxin Joy Jiao, Spencer Papay, Amelia Glaese, John Schulman, and William Fedus. 2024. Measuring short-form factuality in large language models. arXiv preprint arXiv:2411.04368

    2024 arXiv

  25. [33]

    Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. 2025. Browsecomp: A simple yet challenging benchmark for browsing agents. arXiv preprint arXiv:2504.12516

    2025 arXiv

  26. [34]

    Jialong Wu, Baixuan Li, Runnan Fang, Wenbiao Yin, Liwen Zhang, Zhenglin Wang, Zhengwei Tao, Ding-Chu Zhang, Zekun Xi, Xiangru Tang, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. 2025 a . Webdancer: Towards autonomous information seeking agency

    2025

  27. [35]

    Jialong Wu, Wenbiao Yin, Yong Jiang, Zhenglin Wang, Zekun Xi, Runnan Fang, Linhai Zhang, Yulan He, Deyu Zhou, Pengjun Xie, and Fei Huang. 2025 b . https://arxiv.org/abs/2501.07572 Webwalker: Benchmarking llms in web traversal . Preprint, arXiv:2501.07572

    2025 arXiv

  28. [36]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, and 22 others. 2024. Qwen2.5 technical repo...

    2024 arXiv

  29. [37]

    Cohen, Ruslan Salakhutdinov, and Christopher D

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. https://arxiv.org/abs/1809.09600 Hotpotqa: A dataset for diverse, explainable multi-hop question answering . Preprint, arXiv:1809.09600

    2018 arXiv

  30. [38]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629

    2022 arXiv

  31. [39]

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, and 1 others. 2025. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476

    2025 arXiv

  32. [40]

    Tian Yu, Shaolei Zhang, and Yang Feng. 2024. Auto-rag: Autonomous retrieval-augmented generation for large language models. arXiv preprint arXiv:2411.19443

    2024 arXiv

  33. [41]

    Zhenrui Yue, Honglei Zhuang, Aijun Bai, Kai Hui, Rolf Jagerman, Hansi Zeng, Zhen Qin, Dong Wang, Xuanhui Wang, and Michael Bendersky. 2025. https://openreview.net/forum?id=FSjIrOm1vz Inference scaling for long-context retrieval augmented generation . In The Thirteenth Internat...

    2025

  34. [42]

    Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. 2025. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. arXiv preprint arXiv:2504.03160

    2025 arXiv

  35. [43]

    Yuxiang Zheng, Shichao Sun, Lin Qiu, Dongyu Ru, Cheng Jiayang, Xuefeng Li, Jifan Lin, Binjie Wang, Yun Luo, Renjie Pan, and 1 others. 2024. Openresearcher: Unleashing ai for accelerated scientific research. arXiv preprint arXiv:2408.06941

    2024 arXiv

  36. [44]

    Wangchunshu Zhou, Yuchen Eleanor Jiang, Long Li, Jialong Wu, Tiannan Wang, Shi Qiu, Jintian Zhang, Jing Chen, Ruipu Wu, Shuai Wang, and 1 others. 2023. Agents: An open-source framework for autonomous language agents. arXiv preprint arXiv:2309.07870

    2023 arXiv

  37. [45]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  38. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.