ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment

Jiajun Wang; Rui Ye; Siheng Chen; Songhua Liu; Tian Jin; Yijun Lu; Yuwen Du

REVIEW 4 major objections 5 minor 26 references

ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Scoring each search step against clues backtracked from the final answer lets a 4-billion-parameter agent match much larger search systems on browsing benchmarks.

desk verdict ABC is a credible new credit-assignment trick for search-agent training, with solid ablations, but the unvalidated clue set and unmatched baselines keep it from being a home run. read the letter →

arxiv 2608.05102 v1 pith:QAI6DK6L submitted 2026-08-05 cs.AI

Yijun Lu , Rui Ye , Jiajun Wang , Yuwen Du , Tian Jin , Songhua Liu , Siheng Chen This is my paper

classification cs.AI

keywords searchagentscreditassignmentlong-horizonstep-levelrewardsanswerbacktrackingreinforcementlearningsupervisedfine-tuning

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

The paper argues that training long-horizon search agents is held back by trajectory-level supervision, which treats every step in a successful run as good and every step in a failed run as bad. It proposes Answer-Backtracked Credit Assignment (ABC): starting from the known final answer, an LLM actively searches the web backward to recover a chain of intermediate clues, then a scorer rates each step of each training trajectory against those clues, producing a dense per-step reward on a fixed rubric. These step rewards reweight the loss in supervised fine-tuning (ABC-SFT) and replace trajectory-level rewards in GRPO (ABC-GRPO), so a failed trajectory can still reinforce the steps that found real evidence, and a successful trajectory can still penalize steps that discarded it. Training a 4B-parameter agent on 8.5 thousand examples with ABC yields 37.3% on BrowseComp and 39.1% on BrowseComp-ZH (55.3% and 52.9% with context management), the best among same-scale agents and competitive with roughly 30B-parameter systems. The framework turns a single correctness bit per trajectory into fine-grained, answer-anchored supervision for every action.

What carries the argument

The central mechanism is the pair of answer-backtracked anchors and clue-anchored step scoring, which together replace a single outcome bit with a dense per-step scalar. Answer-Backtracked Clue Recovery is an active search loop that starts from the ground-truth answer and travels backward through web evidence to assemble a set of verifiable clues, so each clue is a named entity, fact, attribute, or relation that connects the query constraints to the answer. Clue-Anchored Step Scoring then rates each step of each trajectory against this fixed clue set using an additive rubric on a base score of 1.0, clipped to the interval [0, 2.0]. These step scores are consumed by ABC-SFT, which maps each score through a sigmoid to a per-turn loss weight, and by ABC-GRPO, which uses the scores as step-level rewards with a discounted advantage and the standard clipped GRPO objective. The fixed, answer-backtracked clue set is what keeps the credit signal stable across policy updates and across both successful and failed trajectories.

What would settle it

Train ABSeeker with the same step-scoring pipeline but replace the recovered clue sets with an equal number of random entities and facts drawn from the query text; if this pseudo-clue variant matches the reported benchmark scores, the answer-backtracking component is not the source of the gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that answer-backtracked step-level credit assignment makes long-horizon search agents train better than uniform step treatment, and the evidence is a concrete training pipeline and a set of benchmark results. Given a query and a verified answer, Answer-Backtracked Clue Recovery reconstructs the intermediate evidence chain—entities, facts, attributes, and relations—that a valid search should establish, using the same tool-call protocol as the forward agent to verify each clue against real web content. Clue-Anchored Step Scoring then evaluates every step in every collected trajectory against this fixed clue set, starting from a base score of 1.0 and adding 0.8 for discovering or verifying a correct clue, 0.4 for correctly ruling out a wrong candidate, subtracting 0.8 for incorrectly dismissing a correct clue, and adding or subtracting 1.0 for the correct or wrong final answer, with the total clipped to [0, 2.0]. These scores become per-turn loss weights in ABC-SFT and step-level rewards in ABC-GRPO, and the trained 4B agent ABSeeker outperforms all 4B baselines on five benchmarks while matching several ~30B agents. The paper concludes that the model learns to search more efficiently and more deliberately as a result of this fine-grained credit.

Load-bearing premise

The load-bearing premise is that the recovered clue set is a complete, correct, and fixed record of the evidence a good search must find; if clue recovery misses, invents, or misorders clues, every step reward derived from it is misaligned with true step quality.

Editorial extensions

If this is right

Failed trajectories stop being wasted training data: steps that discover or verify correct clues receive positive credit even when the final answer is wrong, so the agent can learn from near-misses.
Successful trajectories stop being treated as uniformly correct: erroneous or redundant steps are down-weighted or penalized, which should reduce reward hacking and sloppy intermediate reasoning.
The clue set provides a fixed, externally anchored evaluation criterion that does not shift with the policy's own beliefs, avoiding the instability of model-based credit signals.
Because the only requirement is a unique, verifiable final answer, the ABC pipeline transfers to any benchmark or task with that property, and the paper shows gains generalize from BrowseComp-style training to xbench and GAIA.
The longer trajectories observed with ABC-GRPO suggest that step-level credit encourages more extensive exploration and verification rather than premature answer submission.

Reading between the lines

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

An untested risk is that the step scorer receives the correct answer as part of its input, so scores may partly reward steps that merely mention the answer or its clues rather than genuinely verify evidence; a cleaner variant would score steps without revealing the answer.
The recovered clue set is assumed to be complete and correct, but the paper reports no quantitative check of recovery accuracy; a perturbation study that corrupts or shuffles clues would reveal how much of the gain actually comes from the backtracked content versus the mere presence of dense supervision.
The method's dependence on a single verifiable answer suggests it will not transfer directly to open-ended research tasks with multiple defensible answers, where no unique evidence chain exists to anchor the scoring.
A practical extension the paper does not explore is using the recovered clue chains at inference time as a search guide or verification checklist, since they encode the evidence structure the agent is expected to discover.

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.

Desk Editor's Note

read the letter

The headline: ABSeeker proposes a genuinely new way to assign step-level credit in long-horizon search-agent training—answer-backtracked clue recovery plus an LLM rubric scorer—and the controlled ablations support the core claim. This is not a restatement of IGPO, CSO, SAPO, or MindDR; the credit signal here is dense and step-local, and the idea of using the verified answer to reconstruct an evidence chain is simple and transferable. The paper also ships code and a model, which makes the work reproducible.

What it does well: the reward distribution analysis (Figure 4) is a nice sanity check that failed trajectories contain useful steps and successful ones contain duds, which motivates the method. The ablations in Table 3 are the right comparison—ABC-SFT vs standard SFT, ABC-GRPO vs standard GRPO on the same data—and they show consistent gains across four of five benchmarks for SFT and all five for GRPO. That is credible evidence that dense step-level supervision helps, independent of the absolute numbers.

The soft spots are real but not fatal. The recovered clue set is the load-bearing component, and the paper gives no precision/recall numbers or human-verified check of the clue-recovery step. The stress-test note is right that if clue recovery is noisy, step rewards misalign; and because the scorer also sees the correct answer, there's a risk of answer leakage rewarding steps that merely mention answer-related entities. This is a legitimate concern, but it doesn't sink the paper: the ablations against standard SFT/GRPO still show gains, and those baselines don't use the clue signal at all. The correct fix is an auxiliary validation study—sample a few dozen questions, have a human check clue completeness and scoring accuracy—and the paper should do that before publication.

The comparison to 30B agents is overstated. The 55.3% BrowseComp score uses context management, while the 30B baselines are reported without it (the asterisks matter). Without context management, ABSeeker's 37.3% is below most 30B agents. The claim "matches 30B" is true only under a non-equivalent evaluation condition. The paper should either compare like-for-like or soften the claim.

Minor: no error bars or significance tests on the three evaluation runs; the RL training uses only 1,000 questions; and the same DeepSeek-V4-Flash model is used for clue recovery, step scoring, and final-answer evaluation, which is a shared-bias risk that a human-verified subset would mitigate.

Overall, this is a serious piece of work with a novel method, credible ablations, and honest limitations in the conclusion. It deserves peer review. I'd recommend accepting it with a request for clue-validation evidence and matched context-management comparisons. It's the kind of paper I'd bring to a reading group working on agent training.

Referee Report

4 major / 5 minor

Summary. The paper proposes Answer-Backtracked Credit Assignment (ABC) for training long-horizon search agents. Given a query and a verified answer, an LLM ReAct loop recovers a set of intermediate clues C; a second LLM scores every step in a trajectory against C (Eq. 4), producing dense step rewards that reweight SFT losses (Eq. 5) and replace trajectory-level rewards in GRPO (Eqs. 6-7). The authors train ABSeeker, a Qwen3.5-4B agent, on 8.5k OpenSeeker trajectories and report that it outperforms standard SFT/GRPO in ablations and is competitive with 30B-scale search agents on BrowseComp, BrowseComp-ZH, xbench, and GAIA-text.

Significance. ABC is a well-motivated and timely approach to process supervision for long-horizon search, and the idea of rewarding useful steps in failed trajectories while penalizing erroneous steps in successful ones is conceptually clean. The strengths include the explicit separation of clue recovery from step scoring, the open-sourced code and model, and ablation results that, taken at face value, support the directional claim that ABC-SFT and ABC-GRPO improve over uniform SFT and GRPO on most benchmarks. However, the central mechanism rests on an unvalidated clue set and an answer-aware scorer; the significance of the paper is therefore conditional on the additional validation and statistical rigor that the authors should provide.

major comments (4)

[Section 3.2, Eq. (3), and Section 3.3, Eq. (4)] The recovered clue set C is asserted to provide 'reliable, answer-backtracked reference points,' but no precision, recall, or human-validated quality measurement is reported for the clue recovery step. Because every step reward in Eq. (4) is computed against C, and because those rewards enter both the ABC-SFT weights in Eq. (5) and the ABC-GRPO rewards in Eqs. (6)-(7), any systematic omission or invention in C directly misaligns all downstream credit. Please add a human audit of a sample of recovered clue sets and a perturbation test (e.g., dropping, adding, or shuffling clues and measuring the change in step rewards and final accuracy) to demonstrate that the ABC signal is robust to realistic recovery errors.
[Section 3.3 and Appendix C.2] The scoring prompt provides the scorer with the Correct Answer and the full Core Clues list, and criterion 5 awards +1.0 for submitting the correct final answer. This creates a concrete leakage path: a step that merely names entities from the answer or the provided clues can receive credit without genuinely discovering or verifying evidence. Because the same proprietary model family is used for clue recovery, step scoring, and final-answer evaluation, the dense signal may partly reflect the judge's preferences rather than intrinsic step quality. Please add a controlled experiment in which the scorer is blinded to the correct answer or to the clue labels, and compare the resulting reward assignments and downstream performance, to quantify the amount of answer leakage.
[Table 3 and Section 4.2] The core ablation results are reported as single numbers without error bars, confidence intervals, or significance tests. Some differences are small or in the wrong direction (e.g., xbench-2505: 72.0 for ABC-SFT vs 73.0 for standard SFT), while others are modest (BrowseComp: 30.8 vs 28.5). Because the paper claims that ABC 'consistently outperforms' standard SFT/GRPO, please report multiple seeds or per-evaluation variance and a significance test. Without this, the central claim is not fully supported by the evidence as presented.
[Table 2 and Section 4.2] The headline comparison is confounded by context management. ABSeeker's 55.3% and 52.9% results on BrowseComp and BrowseComp-ZH are obtained with 256K context and the discard-all strategy, while several baselines are marked with an asterisk for no context management; it is unclear which baselines used the same context budget and which did not. The paper should explicitly tabulate the context-management setting for every baseline and compare all methods under identical inference budgets before claiming that ABSeeker outperforms same-scale agents or matches 30B agents.

minor comments (5)

[Section 3.4.1 and Appendix A] The loss weight w(r_t) is defined in the main text as a sigmoid with parameters α and β, but Appendix A gives the concrete form w(r_t) = 2σ(2(r_t − 1)). Please reconcile these definitions and state the exact values of α and β used in the experiments.
[Eq. (2) and Appendix C.3] Equation (2) defines the trajectory reward as exact string equality a = a*, while evaluation uses an LLM judge (Appendix C.3). Please clarify whether training rewards use exact matching, LLM-based matching, or both, since this affects the interpretation of the reward signal.
[Appendix A and Section 4.1] The frameworks 'Slime' and 'veRL' are mentioned without references or URLs; please add citations or links so that the training setup is reproducible.
[Section 4.2, Figure 4] The caption of Figure 4 does not explain the x-axis labels (e.g., '0–10%' and '90–100%') clearly enough; please state explicitly that each bar represents one-tenth of the steps within a trajectory, ordered temporally.
[Section 4.1 and Appendix B] The DeepSeek-V4-Flash model is used for clue recovery, step scoring, and final-answer evaluation, but no version or snapshot date is given. Because proprietary models change over time, please specify the exact version and date used for all reported results.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the step-level rewards are rubric-based outputs of an explicit scoring procedure, and the central claim is validated on external benchmarks via ablations.

full rationale

The paper's derivation chain is self-contained. Clue recovery (Eq. 3) maps the query and verified answer to a clue set C via a ReAct LLM loop; Clue-Anchored Step Scoring (Eq. 4 and Table 1) then assigns each step a score from accumulated rubric deltas relative to C. The rewards are not algebraically determined by the final answer: the rubric requires per-step content (reasoning, tool call, tool response) to be matched against the recovered clue set, and the scorer is instructed to name the specific clue/entity for each applied criterion (Appendix C.2). Equations (5)-(7) use these rewards merely to reweight SFT losses and to define GRPO advantages; neither equation collapses to the trajectory-level answer reward of Eq. (2), because intermediate steps can receive positive or negative deltas independent of the final outcome. The claimed gains are established by ablations against standard SFT and standard GRPO (Table 3) on external benchmarks with ground-truth answers, not by the reward construction itself. The citations to OpenSeeker (training data) and LongSeeker (context management) involve overlapping authors but are not load-bearing for the credit-assignment claim: OpenSeeker supplies public trajectories and LongSeeker supplies an orthogonal context-management technique. The main unaddressed risk is that the recovered clue set is not quantitatively validated for completeness or accuracy (Section 3.2 asserts 'Clues that survive this verification serve as reliable...' without precision/recall or human checks), so noisy clue recovery could misalign step rewards; this is an empirical robustness/correctness concern, not a construction-level circularity.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The method rests on an external LLM that both produces and consumes the credit signals: clue recovery defines the reward anchors, step scoring uses those anchors, and the same model family evaluates final answers. The rubric weights and RL hyperparameters are hand-chosen free parameters. No new entities are postulated.

free parameters (8)

base step score = 1.0
Every step starts at 1.0 (Table 1, Eq. 4); chosen by hand to make neutral exploration unpenalized.
clue discovery delta = +0.8
Rubric delta for discovering or verifying a correct clue (Table 1); hand-chosen.
incorrect dismissal delta = -0.8
Rubric penalty for incorrectly dismissing a correct clue (Table 1); hand-chosen.
rule-out delta = +0.4
Rubric reward for ruling out an incorrect candidate (Table 1); hand-chosen.
final answer deltas = +1.0 / -1.0
Rubric deltas for correct and incorrect final submissions (Table 1); hand-chosen.
reward clipping upper bound = 2.0
Step scores clipped to [0,2] in Eq. 4; hand-chosen.
SFT weight mapping = alpha=2, beta=1, w=2 sigma(2(r-1))
Appendix A: reward-to-loss-weight sigmoid; hyperparameters chosen by hand.
RL discount factor = 0.25
gamma in Eq. 7; hyperparameter chosen by hand.

assumptions (4)

domain assumption For each (query, answer) pair there is a single valid evidence path that clue recovery can reconstruct.
Section 3.2 states that constraints plus the verified answer 'implicitly define a single valid evidence path'. If paths are multiple or recovery is incomplete, step scoring is misaligned.
ad hoc to paper The recovered clue set C is complete and correct and can serve as fixed anchors for scoring every step.
Section 3.3 treats C as ground truth for scoring; no quantitative validation of clue completeness is given, and the scorer is shown the correct answer, which may bias scores.
domain assumption The LLM (DeepSeek-V4-Flash) can reliably assign step scores according to the rubric and can reliably judge final-answer correctness.
Appendix C.2 and C.3 rely on DeepSeek-V4-Flash for both rewards and evaluation; systematic judge bias would propagate into training and into all reported benchmark numbers.
ad hoc to paper Step scores in [0,2] are a valid training signal for policy optimization.
Section 3.4 uses r_t directly as loss weights and RL rewards; the paper does not analyze reward noise, calibration, or distribution shift.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment." pith.science (2026). https://pith.science/paper/QAI6DK6L

@misc{pith2026260805102,
  author       = {Pith},
  title        = {Pith review of: ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QAI6DK6L}},
  note         = {Machine review of arXiv:2608.05102}
}

read the original abstract

Long-horizon search agents must make multiple sequential actions (steps) to search, retrieve, verify, and integrate evidence to reach a final answer. However, existing methods for training these agents typically treat all steps within a trajectory uniformly during both supervised fine-tuning (SFT) and reinforcement learning (RL), failing to distinguish useful actions from erroneous or redundant ones. In this paper, we propose Answer-Backtracked Credit Assignment (ABC), a fine-grained credit assignment framework for training long-horizon search agents by converting sparse trajectory-level outcomes into dense step-level supervision that rewards useful actions (even in failed trajectories) while suppressing erroneous or redundant actions. Specifically, given a potentially obscure query and its corresponding ground-truth answer, ABC first performs Answer-Backtracked Clue Recovery, which traces back from the answer to recover intermediate clues required to solve the question. It then applies Clue-Anchored Step Scoring to evaluate each search step against these clues, converting sparse binary outcome supervision into dense step-level rewards. Based on these rewards, we develop ABC-SFT, which reweights the loss of each turn, and ABC-GRPO, which uses the step-level scores as rewards in GRPO. Building on this framework, we train ABSeeker based on Qwen3.5-4B with only 8.5k examples. ABSeeker achieves 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, the scores further improve to 55.3% and 52.9%, respectively, significantly outperforming same-scale (4B) agents and even matching the performance of larger ones (approximately 30B). These results demonstrate the effectiveness of answer-backtracked step-level credit assignment for training long-horizon search agents.

Figures

Figures reproduced from arXiv: 2608.05102 by the authors.

**Figure 2.** Overview of the training pipeline, which consists of two core stages. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** An illustrative example of Answer-Backtracked Clue Recovery and Clue-Anchored Step Scoring. Left: Given a multi-constraint query and the verified answer, the recovery produces six intermediate evidence clues (c1–c6) that form a verified evidence chain connecting the query to the answer. Right: Selected steps from a sampled trajectory are evaluated against the recovered clue set, with rewards reflecting the quality … view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Distribution of step rewards across the 8.5K SFT trajectories. The left half shows successful [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: RL training dynamics. Left: Performance during training. Right: Average number of interaction turns [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 5 canonical work pages

[1]

Seed2.0 model card: Towards intelligence frontier for real-world complexity.arXiv preprint arXiv:2607.00248,

Bytedance Seed. Seed2.0 model card: Towards intelligence frontier for real-world complexity.arXiv preprint arXiv:2607.00248,

arXiv
[3]

xbench: Tracking agents productivity scaling with profession-aligned real-world evaluations.arXiv preprint arXiv:2506.13651, 2025a

Kaiyuan Chen, Yixin Ren, Yang Liu, Xiaobo Hu, Haotong Tian, Tianbao Xie, Fangfu Liu, Haoye Zhang, Hongzhang Liu, Yuan Gong, et al. xbench: Tracking agents productivity scaling with profession-aligned real-world evaluations.arXiv preprint arXiv:2506.13651, 2025a. M Chen, L Sun, T Li, H Sun, Y Zhou, C Zhu, H Wang, JZ Pan, W Zhang, H Chen, et al. Research: L...

arXiv 2025
[4]

Openseeker: Democ- ratizing frontier search agents by fully open-sourcing training data.arXiv preprint arXiv:2603.15594,

Yuwen Du, Rui Ye, Shuo Tang, Xinyu Zhu, Yijun Lu, Yuzhu Cai, and Siheng Chen. Openseeker: Democ- ratizing frontier search agents by fully open-sourcing training data.arXiv preprint arXiv:2603.15594,

arXiv
[5]

Beyond ten turns: Unlocking long-horizon agentic search with large-scale asynchronous rl.arXiv preprint arXiv:2508.07976,

Jiaxuan Gao, Wei Fu, Minyang Xie, Shusheng Xu, Chuyi He, Zhiyu Mei, Banghua Zhu, and Yi Wu. Beyond ten turns: Unlocking long-horizon agentic search with large-scale asynchronous rl.arXiv preprint arXiv:2508.07976,

arXiv
[6]

Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516,

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

arXiv
[8]

Beyond Trajectory Rewards: Step-level Credit Assignment for Agentic Search via Graph Modeling

Yuchen Liu, Yingjie Feng, Lixiong Qin, Jiasi Chen, Jianing Yu, Sheng Gao, Sheng Yang, and Weiran Xu. Beyond trajectory rewards: Step-level credit assignment for agentic search via graph modeling.arXiv preprint arXiv:2605.29697,

work page Pith review arXiv
[9]

Longseeker: Elastic context orchestration for long-horizon search agents.arXiv preprint arXiv:2605.05191,

10 Yijun Lu, Rui Ye, Yuwen Du, Jiajun Wang, Songhua Liu, and Siheng Chen. Longseeker: Elastic context orchestration for long-horizon search agents.arXiv preprint arXiv:2605.05191,

arXiv
[10]

Gaia: a benchmark for general ai assistants

Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants. InInternational Conference on Learning Representations, volume 2024, pages 9025–9049,

2024

Show all 26 references

[11]

Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300,

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

arXiv
[13]

R1-searcher: Incentivizing the search capability in llms via reinforcement learning.arXiv preprint arXiv:2503.05592,

Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning.arXiv preprint arXiv:2503.05592,

arXiv
[14]

Beyond turn limits: Training deep search agents with dynamic context window

Qiaoyu Tang, Hao Xiang, Le Yu, Bowen Yu, Yaojie Lu, Xianpei Han, Le Sun, WenJuan Zhang, Pengbo Wang, Shixuan Liu, et al. Beyond turn limits: Training deep search agents with dynamic context window. arXiv preprint arXiv:2510.08276,
[15]

Mind deepresearch technical report.arXiv preprint arXiv:2604.14518,

MindDR Team and Li Auto Inc. Mind deepresearch technical report.arXiv preprint arXiv:2604.14518,

arXiv
[16]

Mirothinker-1.7 & h1: Towards heavy-duty research agents via verification.arXiv preprint arXiv:2603.15726, 2026a

MiroMind Team, S Bai, L Bing, L Lei, R Li, X Li, X Lin, E Min, L Su, B Wang, et al. Mirothinker-1.7 & h1: Towards heavy-duty research agents via verification.arXiv preprint arXiv:2603.15726, 2026a. Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guang...
[17]

Dr-venus: Towards frontier edge-scale deep research agents with only 10k open data.arXiv preprint arXiv:2604.19859, 2026b

Venus Team, Sunhao Dai, Yong Deng, Jinzhen Lin, Yusheng Song, Guoqing Wang, Xiaofeng Wu, Yuqi Zhou, Shuo Yang, Zhenzhe Ying, et al. Dr-venus: Towards frontier edge-scale deep research agents with only 10k open data.arXiv preprint arXiv:2604.19859, 2026b. Guoqing Wang, Sunhao D...

arXiv
[18]

Browsecomp: A simple yet challenging benchmark for browsing agents.arXiv preprint arXiv:2504.12516,

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

arXiv
[19]

Quest: Training frontier deep research agents with fully synthetic tasks

Jian Xie, Tianhe Lin, Zilu Wang, Yuting Ning, Yuekun Yao, Tianci Xue, Zhehao Zhang, Zhongyang Li, Kai Zhang, Yufan Wu, et al. Quest: Training frontier deep research agents with fully synthetic tasks. arXiv preprint arXiv:2605.24218,

arXiv
[20]

Deepseek-v4: Towards highly efficient million-token context intelligence.arXiv preprint arXiv:2606.19348,

Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, et al. Deepseek-v4: Towards highly efficient million-token context intelligence.arXiv preprint arXiv:2606.19348,
[21]

Qwen3 technical report.arXiv preprint arXiv:2505.09388,

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

arXiv
[22]

React: Synergizing reasoning and acting in language models

Shunyu Yao, Jeffrey Zhao, Dian Yu, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InNeurIPS 2022 F oundation Models for Decision Making Workshop,

2022
[24]

Deepresearcher: Scaling deep research via reinforcement learning in real-world environments

Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processin...

2025
[25]

Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese.arXiv preprint arXiv:2504.19314,

Peilin Zhou, Bruce Leon, Xiang Ying, Can Zhang, Yifan Shao, Qichen Ye, Dading Chong, Zhiling Jin, Chenxuan Xie, Meng Cao, et al. Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese.arXiv preprint arXiv:2504.19314,

arXiv
[26]

We randomly select 8.5K trajectories from OpenSeeker (Du et al., 2026), consisting of 5.5K correct and 3.0K incorrect trajectories

12 A Training Details Supervised Fine-Tuning.We implement SFT using Slime and initialize the model from Qwen3.5-4B. We randomly select 8.5K trajectories from OpenSeeker (Du et al., 2026), consisting of 5.5K correct and 3.0K incorrect trajectories. We train on these 8.5K scored...

2026
[2022]

Glm-5: from vibe coding to agentic engineering.arXiv preprint arXiv:2602.15763,

Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chenghua Huang, Chengxing Xie, et al. Glm-5: from vibe coding to agentic engineering.arXiv preprint arXiv:2602.15763,

arXiv
[2024]

Openai gpt-5 system card.arXiv preprint arXiv:2601.03267,

Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaugh- lin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. Openai gpt-5 system card.arXiv preprint arXiv:2601.03267,

arXiv
[2025]

Verified critical step optimization for llm agents.arXiv preprint arXiv:2602.03412,

Mukai Li, Qingcheng Zeng, Tianqing Fang, Zhenwen Liang, Linfeng Song, Qi Liu, Haitao Mi, and Dong Yu. Verified critical step optimization for llm agents.arXiv preprint arXiv:2602.03412,

arXiv
[2026]

Haotian Chen, Xin Cong, Shengda Fan, Yuyang Fu, Ziqin Gong, Yaxi Lu, Yishan Li, Boye Niu, Chengjun Pan, Zijun Song, et al

doi: 10.48550/arXiv.2607.00248. Haotian Chen, Xin Cong, Shengda Fan, Yuyang Fu, Ziqin Gong, Yaxi Lu, Yishan Li, Boye Niu, Chengjun Pan, Zijun Song, et al. Agentcpm-explore: Realizing long-horizon deep exploration for edge-scale agents. arXiv preprint arXiv:2602.06485,

doi arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Seed2.0 model card: Towards intelligence frontier for real-world complexity.arXiv preprint arXiv:2607.00248,

Bytedance Seed. Seed2.0 model card: Towards intelligence frontier for real-world complexity.arXiv preprint arXiv:2607.00248,

arXiv

[2] [3]

xbench: Tracking agents productivity scaling with profession-aligned real-world evaluations.arXiv preprint arXiv:2506.13651, 2025a

Kaiyuan Chen, Yixin Ren, Yang Liu, Xiaobo Hu, Haotong Tian, Tianbao Xie, Fangfu Liu, Haoye Zhang, Hongzhang Liu, Yuan Gong, et al. xbench: Tracking agents productivity scaling with profession-aligned real-world evaluations.arXiv preprint arXiv:2506.13651, 2025a. M Chen, L Sun, T Li, H Sun, Y Zhou, C Zhu, H Wang, JZ Pan, W Zhang, H Chen, et al. Research: L...

arXiv 2025

[3] [4]

Openseeker: Democ- ratizing frontier search agents by fully open-sourcing training data.arXiv preprint arXiv:2603.15594,

Yuwen Du, Rui Ye, Shuo Tang, Xinyu Zhu, Yijun Lu, Yuzhu Cai, and Siheng Chen. Openseeker: Democ- ratizing frontier search agents by fully open-sourcing training data.arXiv preprint arXiv:2603.15594,

arXiv

[4] [5]

Beyond ten turns: Unlocking long-horizon agentic search with large-scale asynchronous rl.arXiv preprint arXiv:2508.07976,

Jiaxuan Gao, Wei Fu, Minyang Xie, Shusheng Xu, Chuyi He, Zhiyu Mei, Banghua Zhu, and Yi Wu. Beyond ten turns: Unlocking long-horizon agentic search with large-scale asynchronous rl.arXiv preprint arXiv:2508.07976,

arXiv

[5] [6]

Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516,

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

arXiv

[6] [8]

Beyond Trajectory Rewards: Step-level Credit Assignment for Agentic Search via Graph Modeling

Yuchen Liu, Yingjie Feng, Lixiong Qin, Jiasi Chen, Jianing Yu, Sheng Gao, Sheng Yang, and Weiran Xu. Beyond trajectory rewards: Step-level credit assignment for agentic search via graph modeling.arXiv preprint arXiv:2605.29697,

work page Pith review arXiv

[7] [9]

Longseeker: Elastic context orchestration for long-horizon search agents.arXiv preprint arXiv:2605.05191,

10 Yijun Lu, Rui Ye, Yuwen Du, Jiajun Wang, Songhua Liu, and Siheng Chen. Longseeker: Elastic context orchestration for long-horizon search agents.arXiv preprint arXiv:2605.05191,

arXiv

[8] [10]

Gaia: a benchmark for general ai assistants

Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants. InInternational Conference on Learning Representations, volume 2024, pages 9025–9049,

2024