Pith. sign in

REVIEW 4 major objections 5 minor 72 references

The paper claims that verification-filtered synthetic tasks, not model size, are the key to training competitive long-horizon search agents.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 03:41 UTC pith:X55CJQS2

load-bearing objection A well-built systems report whose internal training gains are plausible, but the 'frontier-matching' headline rests on heterogeneous benchmark provenance and should not be taken at face value. the 4 major comments →

arxiv 2607.24850 v1 pith:X55CJQS2 submitted 2026-07-25 cs.IR cs.LG

SearchArt: Training Long-Horizon Search Agent with Scalable Synthetic and Verified Task

Lang Mei , Xiaohan Yu , Chong Chen , Liyan Liu , Xiangnan Chen , Jinchao Ma , Chao Feng , Li Huang
show 21 more authors
Siyu Mo Sichen Kang Yunkun Xu Zhihan Yang Zhujun Xue Jingren Zhang Qing He Yingdi Huang Hao Jiang Ziao Ma Zewei Pan Minhao Sun Zhuo Tao Jinzhao Xiao Gangtao Xin Huanyao Zhang Wenjian Zhang Jiangshan Zhang Guojie Zhu Jiaxin Mao Wentao Zhang
This is my paper
classification cs.IR cs.LG
keywords long-horizon search agentssynthetic data verificationdeep searchdeep researchreinforcement learningevidence graphstask synthesisagentic search
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that the bottleneck in training long-horizon search agents is not base-model size but the availability of verified, hard, long-horizon training tasks. It synthesizes large datasets of search and research questions from web documents and automatically built evidence graphs, then filters them through a multi-stage verification pipeline that checks question-answer consistency, trajectory quality, and evidence relevance. The verified trajectories are used for supervised fine-tuning followed by reinforcement learning. With only 27 billion parameters, the resulting agent scores 74.39 on BrowseComp-ZH, 70.06 on BrowseComp, and 52.55 on DeepResearch-Bench, matching or beating substantially larger open and closed systems. If the comparison holds, the implication is that scalable synthetic supervision, not scale alone, is a viable route to competitive search agents.

Core claim

SearchArt's central empirical claim is that a 27B-parameter model, post-trained on verification-filtered synthetic search tasks, reaches 74.39 on BrowseComp-ZH, 70.06 on BrowseComp, 63.49 on BrowseComp-Plus, 64.0 on WideSearch, and 52.55 on DeepResearch-Bench, matching or surpassing frontier closed-source agents on both deep-search and deep-research benchmarks. The framework synthesizes QA pairs from web-derived knowledge graphs and high-quality long-form documents, with difficulty controls (multi-hop evidence chains, evidence source dispersion, clue obfuscation) and a multi-stage verification pipeline that rejects tasks solvable by parametric knowledge, filters out unverifiable or ambiguous

What carries the argument

The load-bearing mechanism is the closed loop of evidence-graph-grounded task synthesis and multi-stage verification. A web-exploring agent expands seed entities into a knowledge graph; subgraphs with high structural complexity and evidence source dispersion are converted into QA tasks with obfuscated clues; and tasks must pass three gates: a baseline model cannot answer from parametric knowledge, a baseline agent with tools can sometimes solve them, and a strong agent's successful long-horizon trajectories are selected for quality. This pipeline converts open-web content into dense, verifiable supervision for intermediate search behavior, and is paired with a composite reward whose turn-lev

Load-bearing premise

The headline comparison assumes that the benchmark scores of rival systems were obtained under equivalent answer-matching and context-management protocols; the paper's own numbers are self-run while most rivals are cited from official or third-party reports that used different protocols.

What would settle it

A third-party re-run of SearchArt and its leading rivals on BrowseComp and BrowseComp-ZH in one shared harness with identical answer-matching, context-discard, and temperature settings; if the re-run collapses the reported gaps (for example, DeepSeek-V3.2's BrowseComp score depends on protocol, 51.4 vs 67.6), the matching-frontier claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A 27B open-weight agent can match systems an order of magnitude larger on deep-search and deep-research benchmarks, so base-model scale is not the sole driver of search-agent capability.
  • The largest gains are on Chinese-language deep search (+12.29 on BrowseComp-ZH) and research-bench tasks (+7.85), suggesting verification-driven synthesis transfers across language and task style.
  • Post-training shifts search behavior: the agent sustains longer trajectories on hard questions (90th percentile turns rise from 81 to 98.8 on BrowseComp) rather than terminating early.
  • Training dynamics show tool-call counts drop then stabilize while reward rises, indicating more efficient search policies, not just more search.
  • The synthesis pipeline removes shortcuts (tasks answerable from parametric memory or a single lookup), so trained agents must gather dispersed evidence rather than rely on memorization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If benchmark comparability is confirmed by shared-harness re-runs, the approach should generalize: any strong base model should gain from verified synthetic long-horizon tasks without any architectural change.
  • The verification pipeline's reliance on a 'strong' teacher agent for trajectory synthesis means the ceiling may be set by teacher capability; a testable extension is to measure how performance scales with teacher strength.
  • The paper's multimodal and AutoResearch extensions are future work; the evidence-graph grounding should transfer to multimodal evidence with cross-modal verification, but this is untested.
  • The main risk to the headline numbers is protocol mismatch between self-run and official-reported scores, so the 'matches frontier agents' claim should be read as conditional on comparable evaluation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents SearchArt, a framework for training long-horizon web-search agents via verification-driven synthetic task generation and a multi-stage post-training pipeline (SFT followed by DAPO RL). The authors synthesize DeepSearch, DeepResearch, and user-oriented QA tasks from web documents and knowledge graphs, filter them through a multi-stage verification pipeline, and train a Qwen3.5-27B model. They report large before/after improvements on five benchmarks (e.g., BrowseComp 61.0 to 70.06, BrowseComp-ZH 62.10 to 74.39, DeepResearch-Bench 44.70 to 52.55) and claim that the 27B model matches or surpasses several frontier closed-source agents on BrowseComp, BrowseComp-ZH, and DeepResearch-Bench.

Significance. If validated, the SearchArt pipeline is a substantial practical contribution: it provides a scalable and detailed recipe for synthetic search-task generation, verification, and post-training, with careful attention to trajectory quality and intermediate tool-use behavior. The training-dynamics analysis (Fig. 7-9) and turn-distribution comparisons (Fig. 8) are informative and indicate that post-training meaningfully changes search persistence. The strongest internal evidence is the before/after comparison on benchmarks where both the base and post-trained models were evaluated with the same self-run harness (BrowseComp-Plus and DeepResearch-Bench). However, the headline claim of matching/surpassing frontier closed-source agents is not yet supported because the comparison mixes evaluation protocols and provenance classes. The paper's contributions would be more solid with a shared-harness evaluation or an explicit sensitivity analysis across protocol choices.

major comments (4)
  1. [Abstract; §5.1; Table 1] The central claim 'matching or surpassing frontier closed-source agents' is not established by the evidence as presented. Table 1 mixes three provenance classes: SearchArt-27B scores are all self-run (*), while most competitor scores are official (†) or third-party (#) numbers obtained under different scaffolds and context-management strategies. Section 5.1 states that SearchArt uses discard-all on BrowseComp, the official scaffold on WideSearch, and a custom dual-space planner/writer on DeepResearch-Bench, with additional SearchArt-specific tools (CHECK gate, context resets, Serper/Jina, goal-conditioned summarization). The table itself shows that protocol changes can move scores by more than 15 points (DeepSeek-V3.2 BrowseComp: 51.4 vs 67.6; GLM-5: 62.0 vs 75.9). Therefore the ranking in Figure 1 is not a controlled comparison; a small protocol difference could change whether SearchArt
  2. [§5.2.2; Table 1] The before/after post-training comparison also mixes provenance. For BrowseComp, BrowseComp-ZH, and WideSearch, the base Qwen3.5-27B scores are marked † (official reports), while the post-trained SearchArt-27B scores are marked * (self-run). Only BrowseComp-Plus and DeepResearch-Bench have self-run base scores. Since the harness, tools, and answer-matching protocol differ between the two columns, part of the reported gains (+9.06, +12.29, +2.9) could reflect the evaluation change rather than the post-training itself. The paper should re-run the base model in the same harness as the post-trained model for all five benchmarks, so the training effect is isolated from protocol effects.
  3. [Appendix B; §4.1] Answer-matching differences further undermine cross-benchmark comparability. SearchArt uses its own LLM judge with custom final-answer extraction and strict numeric equality (e.g., 3.14 vs 3.14159 is a non-match; the judge is instructed not to question the ground truth). Official benchmark scripts may use different extraction and equivalence rules. This is not a problem for same-harness before/after comparisons, but it is a serious confound when comparing SearchArt's self-run scores with official or third-party scores that may use looser or stricter matching. The paper should either adopt the official evaluation scripts for all models, or report a sensitivity analysis using alternative answer-matching protocols.
  4. [§2 (data synthesis); §5] The paper does not report a decontamination analysis between the synthesized training data and the evaluation benchmarks. The synthesis pipeline draws from web documents and generates QA pairs that can resemble benchmark-style multi-hop questions; without measuring n-gram or semantic overlap with BrowseComp, BrowseComp-ZH, WideSearch, or DeepResearch-Bench, the reported scores may be inflated by train/eval leakage. I request an explicit decontamination analysis (e.g., overlap statistics and an ablation trained on the non-overlapping subset) to support the claim that the improvements reflect genuine search-agent capability rather than memorization of similar tasks.
minor comments (5)
  1. [Throughout] The text contains many formatting errors: 'T o' for 'To', '⚶' artifacts, 'model ns' for 'model's', 'inital' for 'initial', and inconsistent hyphenation of benchmark names ('Wide-Search' vs 'WideSearch', 'DeepResearch-Bench' vs 'DeepResearch Bench'). These should be cleaned up.
  2. [§3.1] The caching mechanism description is duplicated almost verbatim in two consecutive paragraphs. Please keep only one version.
  3. [§5.2.1] The text and Table 1 report some numbers inconsistently: e.g., GLM 4.7 DeepResearch-Bench is 50.43 in the text but 50.43 in Table 1 (fine), but MiniMax 2.5 WideSearch appears as 70.3†/53.0∗ in the table while the text mentions only '70.3†/ 53.0∗' without clarification. Also, some rows in Table 1 have multiple protocol variants without a clear explanation of which one is used in Figure 1. Please add a column or footnote specifying the protocol for each number.
  4. [§2.4.4] Stage-III verification uses 'a stronger LLM' and 'the strong agent' without specifying which model(s) were used. Since the choice of verifier can affect which instances are retained, the paper should name the verifier and report its pass rates for reproducibility.
  5. [§4.2.2] The 'discard-all' context management is said to 'reset the context by discarding all the previous tool call history and restart the agent rollout.' The exact trigger threshold ('a predefined fraction of the maximum budget') is not given. Please state the threshold and the restart behavior precisely.

Circularity Check

0 steps flagged

No circular derivation: SearchArt's scores are external-benchmark outcomes, and no fitted parameter or self-citation is shown to force them.

full rationale

The paper's central claim is that a verification-driven synthetic-data pipeline followed by SFT and DAPO RL improves a Qwen3.5-27B agent on five external benchmarks. I checked the derivation chain: synthetic QA generation (§2.1–2.3), multi-stage verification (§2.4), SFT (§3.2), RL with composite rewards (§3.3–3.4), and harness design (§4). None of these stages fits a parameter to, or defines a prediction in terms of, the benchmark scores reported in §5.2.1. The difficulty ranking (Eq. 1–2) and reward formulas (Eq. 5–10) are data-construction and training objectives, not functions of BrowseComp/DeepResearch-Bench outcomes. The only evaluator-like component shared between training and inference is an LLM judge/verifier (Appendix B, §3.4.1), but benchmark numbers are not claimed to be derived from that judge; the paper states official scaffolds/scripts are used for WideSearch and external scores are sourced from official/third-party reports (Table 1 footnotes). Self-citations in related work (e.g., refs. [28–31], [33–34], [51]) are contextual and not load-bearing for the headline result. The mixed-provenance comparison in Table 1 (∗/†/#) raises a legitimate measurement-comparability concern about the 'matching/surpassing frontier' headline, but that is a correctness/risk issue, not circularity; it does not make any reported score equivalent to the paper's inputs by construction. No circular step can be exhibited with a quote, so under the hard-evidence rule the score is 0.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claim mainly rests on domain assumptions about data quality and evaluation comparability rather than mathematical axioms. The free parameters cluster in the reward design (Eq. 5, 10) and curation thresholds; λ1/λ2 are never stated. No invented scientific entities are postulated: the CHECK tool and dual-space hierarchical memory (§4) are software components, not entities requiring independent falsifiable evidence.

free parameters (5)
  • Turn-reward thresholds τ1-τ4 = 0.1, 0.2, 0.7, 0.8
    Set 'empirically' (§3.4.2, §5.1). They define the optimal-plateau shape of R_turn, directly shaping the trajectory-length behavior reported in Fig. 8 and the claimed 'efficient tool-use' learning in Fig. 9.
  • Reward weights λ1, λ2
    Eq. (5) defines R = R_outcome + λ1 R_format + λ2 R_turn but the paper never states λ1 or λ2 anywhere; the composite reward is underspecified. Values presumably tuned but unreported.
  • Difficulty-ranking composite weights
    §2.1.3 says candidate subgraphs are ranked by a 'final difficulty score' integrating structural complexity and ESD, but the weighting of degree/depth/width/ESD is never specified.
  • Verification pipeline thresholds = N attempts; retain n ≤ N/2 (Stage II); 3-100 turns; ≤20% failed tool calls
    §2.4.2-2.4.3 and §3.2: curation thresholds (sampling counts, retention rules, turn and tool-failure cutoffs) are chosen by hand and determine both dataset composition and the reported difficulty statistics.
  • CHECK confidence gate = 90/100
    §4.1: answer finalization requires a verification confidence score above 90; this hand-chosen gate affects final-answer behavior on benchmarks.
axioms (6)
  • domain assumption Third-party benchmark scores cited as †/# are accurate as reported by their original sources.
    §5.1, Table 1: the comparison claim depends on ~40 external numbers the authors did not generate; any inconsistency or favorable selection changes the headline result.
  • domain assumption Teacher trajectories (Qwen3.5-397B, Qwen3.6-Plus, Qwen3.7-max, DeepSeek-V4-Pro judge) are correct enough to serve as supervision.
    §3.2: the SFT corpus is teacher-generated and judge-filtered; errors in either propagate into the policy.
  • domain assumption Synthesized training data do not overlap benchmark questions in a way that inflates scores.
    §2.1-2.2: data are synthesized from web documents and arXiv surveys 2020+; BrowseComp and DeepResearch-Bench are also web/research grounded. No contamination or overlap analysis is reported.
  • domain assumption The same answer-matching judge paradigm used in training rewards is not a source of evaluation inflation.
    §3.4.1 and Appendix B: rule-first, judge-fallback matching is used for rewards; benchmark answer matching follows a similar semantic-matching prompt. If the judge family aligns with the trained model's output style, scores may overstate correctness.
  • standard math DAPO RL objective (Eq. 3-4) and advantage normalization behave as in the cited source.
    §3.3.1: the training loop assumes standard convergence behavior of policy-gradient optimization; not re-derived.
  • domain assumption arXiv title-filter plus LLM quality filtering correctly identifies high-quality reference answers.
    §2.2.1: DeepResearch QA treats surveys as gold answers; wrong or low-quality documents produce misaligned question-answer pairs.

pith-pipeline@v1.3.0-alltime-deepseek · 36693 in / 22842 out tokens · 218920 ms · 2026-08-01T03:41:11.121567+00:00 · methodology

0 comments
read the original abstract

Recent advances in large language models (LLMs) have enabled search agents to autonomously tackle complex tasks across extended search and reasoning horizons. However, training effective search agents remains challenging due to the lack of scalable and long-horizon tasks, and the difficulty of evaluating and correcting intermediate reasoning and tool-use behaviors. We introduce SearchArt, a scalable framework for training long-horizon search agents through verification-driven task synthesis and a multi-stage post-training pipeline. SearchArt constructs large-scale datasets for complex search-, research- and user-oriented tasks by synthesizing diverse information-seeking QA pairs and corresponding search trajectories from web documents and automatically generated evidence graphs. To ensure the reliability of the synthesized data, we design a verification pipeline that jointly evaluates QA consistency, trajectory quality, and the relevance of retrieved evidence. The verified trajectories are subsequently used in a multi-stage training process comprising supervised fine-tuning and reinforcement learning-based policy optimization. Search agents trained with SearchArt exhibit adaptive search planning, iterative evidence aggregation, and complex reasoning over extended interaction horizons. Experimental results demonstrate that, with only (Qwen3.5-) 27B parameters, SearchArt scores 74.39 on BrowseComp-ZH, 70.06 on BrowseComp, and 52.55 on Deepresearch-bench, matching or surpassing frontier closed-source agents on both deepsearch and deepresearch benchmarks.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

72 extracted references · 26 linked inside Pith

  1. [1]

    URL https://huggingface.co/MiniMaxAI/MiniMax-M2.5

    Minimaxai, 2026. URL https://huggingface.co/MiniMaxAI/MiniMax-M2.5

    2026

  2. [2]

    URL https://www.kimi.com/blog/

    Kimi, 2026. URL https://www.kimi.com/blog/

    2026

  3. [3]

    URL https://qwen.ai/blog?id=qwen3.5

    Qwen, 2026. URL https://qwen.ai/blog?id=qwen3.5

    2026

  4. [4]

    Gpt-4 technical report

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

    Pith/arXiv arXiv 2023

  5. [5]

    Claude opus, 2026

    Anthropic. Claude opus, 2026. URL https://claude.ai/

    2026

  6. [6]

    Qwen technical report

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Mao-arag: Multi-agent orchestration for adaptive retrieval-augmented generation

    Yiqun Chen, Erhan Zhang, Lingyong Yan, Shuaiqiang Wang, Jizhou Huang, Dawei Yin, and Jiaxin Mao. Mao-arag: Multi-agent orchestration for adaptive retrieval-augmented generation. arXiv preprint arXiv:2508.01005, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Improving retrieval-augmented generation through multi-agent reinforcement learning

    Yiqun Chen, Lingyong Yan, Weiwei Sun, Xinyu Ma, Yi Zhang, Shuaiqiang Wang, Dawei Yin, Yiming Yang, and Jiaxin Mao. Improving retrieval-augmented generation through multi-agent reinforcement learning. Advances in Neural Information Processing Systems, 38:121336–121367, 2026. ∗Team Leader 34

    2026

  9. [9]

    Beyond monolithic architectures: A multi-agent search and knowledge optimization framework for agentic search

    Yiqun Chen, Lingyong Yan, Zixuan Yang, Erhan Zhang, Jiashu Zhao, Shuaiqiang Wang, Dawei Yin, and Jiaxin Mao. Beyond monolithic architectures: A multi-agent search and knowledge optimization framework for agentic search. arXiv preprint arXiv:2601.04703, 2026

    arXiv 2026

  10. [10]

    Unitymas-o: A general rl optimization framework for llm-based multi-agent systems

    Yiqun Chen, Wei Yang, Erhan Zhang, Shijie Wang, Qi Liu, Zechun Niu, Bin Zhang, Haitao Li, Rui Li, Lingyong Yan, et al. Unitymas-o: A general rl optimization framework for llm-based multi-agent systems. arXiv preprint arXiv:2605.26646, 2026

    Pith/arXiv arXiv 2026

  11. [11]

    Jade: Bridging the strategic-operational gap in dynamic agentic rag

    Yiqun Chen, Erhan Zhang, Tianyi Hu, Shijie Wang, Zixuan Yang, Meizhi Zhong, Xiaochi Wei, Yan Gao, Yi Wu, Yao Hu, et al. Jade: Bridging the strategic-operational gap in dynamic agentic rag. arXiv preprint arXiv:2601.21916, 2026

    arXiv 2026

  12. [12]

    Browsecomp-plus: A more fair and transparent evaluation benchmark of deep-research agent

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

    Pith/arXiv arXiv 2025

  13. [13]

    Redsearcher: A scalable and cost-efficient framework for long-horizon search agents

    Zheng Chu, Xiao Wang, Jack Hong, Huiming Fan, Yuqi Huang, Yue Yang, Guohai Xu, Chenxiao Zhao, Cheng Xiang, Shengchao Hu, et al. Redsearcher: A scalable and cost-efficient framework for long-horizon search agents. arXiv preprint arXiv:2602.14234, 2026

    arXiv 2026

  14. [14]

    Fort-searcher: Synthesizing shortcut-resistant search tasks for training deep search agents

    Jia Deng, Yimeng Chen, Xiaoqing Xiang, Ziyang Zeng, Shuo Tang, Wayne Xin Zhao, Feng Chang, Chuan Hao, Yuan Wei, Ran Tao, et al. Fort-searcher: Synthesizing shortcut-resistant search tasks for training deep search agents. arXiv preprint arXiv:2606.12087, 2026

    Pith/arXiv arXiv 2026

  15. [15]

    Deepresearch bench: A comprehensive benchmark for deep research agents

    Mingxuan Du, Benfeng Xu, Chiwei Zhu, Xiaorui Wang, and Zhendong Mao. Deepresearch bench: A comprehensive benchmark for deep research agents. arXiv preprint arXiv:2506.11763, 2025

    Pith/arXiv arXiv 2025

  16. [16]

    Openseeker: Democratizing frontier search agents by fully open-sourcing training data

    Yuwen Du, Rui Ye, Shuo Tang, Xinyu Zhu, Yijun Lu, Yuzhu Cai, and Siheng Chen. Openseeker: Democratizing frontier search agents by fully open-sourcing training data. arXiv preprint arXiv:2603.15594, 2026

    arXiv 2026

  17. [17]

    Retrieval-augmented generation for large language models: A survey

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2(1), 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Cosearchagent: a lightweight collaborative search agent with large lan- guage models

    Peiyuan Gong, Jiamian Li, and Jiaxin Mao. Cosearchagent: a lightweight collaborative search agent with large lan- guage models. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2729–2733, 2024

    2024

  19. [19]

    Cardrewriter: Leveraging knowledge cards for long-tail query rewriting on short-video platforms

    Peiyuan Gong, Feiran Zhu, Yaqi Yin, Chenglei Dai, Chao Zhang, Kai Zheng, Wentian Bao, Jiaxin Mao, and Yi Zhang. Cardrewriter: Leveraging knowledge cards for long-tail query rewriting on short-video platforms. In Proceedings of the ACM Web Conference 2026, pages 8029–8040, 2026

    2026

  20. [20]

    Gemini, 2026

    Google. Gemini, 2026. URL https://gemini.google.com

    2026

  21. [21]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Step-deepresearch technical report

    Chen Hu, Haikuo Du, Heng Wang, Lin Lin, Mingrui Chen, Peng Liu, Ruihang Miao, Tianchi Yue, Wang You, Wei Ji, et al. Step-deepresearch technical report. arXiv preprint arXiv:2512.20491, 2025

    arXiv 2025

  23. [23]

    Search-r1: Training llms to reason and leverage search engines with reinforcement learning

    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, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems , 33:9459–9474, 2020

    2020

  25. [25]

    Webthinker: Empowering large reasoning models with deep research capability

    Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yongkang Wu, Ji-Rong Wen, Yutao Zhu, and Zhicheng Dou. Webthinker: Empowering large reasoning models with deep research capability. Advances in Neural Information Processing Systems, 38:120091–120131, 2026

    2026

  26. [26]

    Deepseek-v3 technical report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Llm4ranking: An easy-to-use framework of utilizing large language models for document reranking

    Qi Liu, Haozhe Duan, Yiqun Chen, Quanfeng Lu, Weiwei Sun, and Jiaxin Mao. Llm4ranking: An easy-to-use framework of utilizing large language models for document reranking. arXiv preprint arXiv:2504.07439, 2025. 35

    Pith/arXiv arXiv 2025

  28. [28]

    Improving first-stage retrieval of point- of-interest search by pre-training models

    Lang Mei, Jiaxin Mao, Juan Hu, Naiqiang Tan, Hua Chai, and Ji-Rong Wen. Improving first-stage retrieval of point- of-interest search by pre-training models. ACM T ransactions on Information Systems, 42(3):1–27, 2023

    2023

  29. [29]

    Dense retrieval for aggregated search

    Lang Mei, Sijie Liu, Ziyuan Zhao, Rolan Yan, Jiaxin Mao, and Ji-rong Wen. Dense retrieval for aggregated search. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management , pages 2095–2104, 2025

    2095

  30. [30]

    A survey of multimodal retrieval-augmented generation

    Lang Mei, Siyu Mo, Zhihan Yang, and Chong Chen. A survey of multimodal retrieval-augmented generation. arXiv preprint arXiv:2504.08748, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Ai-searchplanner: Modular agentic search via pareto-optimal multi-objective reinforcement learning

    Lang Mei, Zhihan Yang, Xiaohan Yu, Huanyao Zhang, and Chong Chen. Ai-searchplanner: Modular agentic search via pareto-optimal multi-objective reinforcement learning. arXiv preprint arXiv:2508.20368, 2025

    arXiv 2025

  32. [32]

    A comprehensive overview of large language models

    Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. ACM T ransactions on Intelligent Systems and T echnology, 16(5):1–72, 2025

    2025

  33. [33]

    Distributionally robust optimization for unbiased learning to rank

    Zechun Niu, Lang Mei, Chong Chen, and Jiaxin Mao. Distributionally robust optimization for unbiased learning to rank. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2266–2275, 2025

    2025

  34. [34]

    Addressing personalized bias for unbiased learning to rank

    Zechun Niu, Lang Mei, Liu Yang, Ziyuan Zhao, Qiang Yan, Jiaxin Mao, and Ji-Rong Wen. Addressing personalized bias for unbiased learning to rank. InProceedings of the 34th ACM International Conference on Information and Knowledge Management, pages 2222–2231, 2025

    2025

  35. [35]

    Gpt, 2026

    OpenAI. Gpt, 2026. URL https://chatgpt.com

    2026

  36. [36]

    Bytedance Seed. Seed2. 0 model card: T owards intelligence frontier for real-world complexity. arXiv preprint arXiv:2607.00248, 2026

    Pith/arXiv arXiv 2026

  37. [37]

    Agentic reasoning and tool integration for llms via reinforcement learning

    Joykirat Singh, Raghav Magazine, Yash Pandya, and Akshay Nambi. Agentic reasoning and tool integration for llms via reinforcement learning. arXiv preprint arXiv:2505.01441, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    R1-searcher: Incentivizing the search capability in llms via reinforcement learning

    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, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Mirothinker-1.7 & h1: T owards heavy-duty research agents via verification

    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: T owards heavy-duty research agents via verification. arXiv preprint arXiv:2603.15726, 2026

    arXiv 2026

  40. [40]

    T ongyi deepresearch technical report

    T ongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, et al. T ongyi deepresearch technical report. arXiv preprint arXiv:2510.24701, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    Browsecomp: A simple yet challenging benchmark for browsing agents

    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, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Widesearch: Benchmarking agentic broad info-seeking

    Ryan Wong, Jiawei Wang, Junjie Zhao, Li Chen, Yan Gao, Long Zhang, Xuan Zhou, Zuo Wang, Kai Xiang, Ge Zhang, et al. Widesearch: Benchmarking agentic broad info-seeking. arXiv preprint arXiv:2508.07999, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Agentic reasoning: Reasoning llms with tools for the deep research

    Junde Wu, Jiayuan Zhu, and Yuyuan Liu. Agentic reasoning: Reasoning llms with tools for the deep research. arXiv preprint arXiv:2502.04644, 9, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Agentic reasoning: A streamlined framework for en- hancing llm reasoning with agentic tools

    Junde Wu, Jiayuan Zhu, Yuyuan Liu, Min Xu, and Yueming Jin. Agentic reasoning: A streamlined framework for en- hancing llm reasoning with agentic tools. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 28489–28503, 2025

    2025

  45. [45]

    Search-p1: Path-centric reward shaping for stable and efficient agentic rag training

    Tianle Xia, Ming Xu, Lingxiang Hu, Yiding Sun, Wenwei Li, Linfang Shang, Liqun Liu, Peng Shu, Huan Yu, and Jie Jiang. Search-p1: Path-centric reward shaping for stable and efficient agentic rag training. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) , pages 517–534, 2026

    2026

  46. [46]

    Deepseek-v4: T owards highly efficient million-token context intelligence.arXiv preprint arXiv:2606.19348, 2026

    Anyi Xu, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chenchen Ling, et al. Deepseek-v4: T owards highly efficient million-token context intelligence.arXiv preprint arXiv:2606.19348, 2026

    arXiv 2026

  47. [47]

    T ournament-grpo: Group-wise tournament rewards for reinforcement learning in open-ended long-form generation

    Zixuan Yang, Yiqun Chen, Wei Yang, Erhan Zhang, Zihan Shen, Xiaochi Wei, Yan Gao, Yi Wu, Yao Hu, and Jiaxin Mao. T ournament-grpo: Group-wise tournament rewards for reinforcement learning in open-ended long-form generation. arXiv preprint arXiv:2605.26958, 2026. 36

    Pith/arXiv arXiv 2026

  48. [48]

    React: Synergizing reasoning and acting in language models

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

    Pith/arXiv arXiv 2022

  49. [49]

    A survey on multimodal large language models

    Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, T ong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12):nwae403, 2024

    2024

  50. [50]

    Dapo: An open-source llm reinforcement learning system at scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale. Advances in Neural Information Processing Systems, 38:113222–113244, 2026

    2026

  51. [51]

    M3 searcher: Modular multimodal information seeking agency with retrieval-oriented reasoning

    Xiaohan Yu, Chao Feng, Lang Mei, and Chong Chen. M3 searcher: Modular multimodal information seeking agency with retrieval-oriented reasoning. arXiv preprint arXiv:2601.09278, 2026

    arXiv 2026

  52. [52]

    Glm-5: from vibe coding to agentic engineering

    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, 2026

    Pith/arXiv arXiv 2026

  53. [53]

    Oases: Outcome-aligned search-evaluation co-training for agentic search

    Erhan Zhang, Yiqun Chen, Zechun Niu, Wei Yang, Xiaochi Wei, Yan Gao, Yi Wu, Yao Hu, and Jiaxin Mao. Oases: Outcome-aligned search-evaluation co-training for agentic search. arXiv preprint arXiv:2604.03675, 2026

    Pith/arXiv arXiv 2026

  54. [54]

    A survey of large language models

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 1(2), 2023

    Pith/arXiv arXiv 2023

  55. [55]

    Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese

    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, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    Marco deepresearch: Unlocking efficient deep research agents via verification-centric design

    Bin Zhu, Qianghuai Jia, Tian Lan, Junyang Ren, Feng Gu, Feihu Jiang, Longyue Wang, Zhao Xu, and Weihua Luo. Marco deepresearch: Unlocking efficient deep research agents via verification-centric design. arXiv preprint arXiv:2603.28376, 2026. 37 A Case A.1 DeepSearch QA English Example (125-turn tool use with Qwen3.7-Max) Question: The British telegraph dir...

    arXiv 2026

  57. [57]

    [citation]1, 4[/citation] • It is expected to exceed 300 EFLOPS in 2025 and 2,500 EFLOPS by 2030

    Overview of China ns Computing Power Market 1.1 Market Size and Growth [citation]1, 2, 3, 4, 9, 18[/citation] 1.1.1 Total Computing Power Capacity • Chinans total computing power capacity reached 280 EFLOPS in 2024. [citation]1, 4[/citation] • It is expected to exceed 300 EFLOPS in 2025 and 2,500 EFLOPS by 2030. [citation]2, 9[/citation] • Chinans computi...

    2024

  58. [58]

    [citation]13[/citation] • The share of inference demand is expected to reach 72.6% in 2027 and 73.0% in 2028

    Structural Characteristics of the Computing Power Market 2.1 Computing Power Mix [citation]1, 3, 13[/citation] 2.1.1 Shares of Training and Inference Computing • In 2023, training workloads accounted for 58.7% of computing power demand, while inference workloads accounted for 41.3%. [citation]13[/citation] • The share of inference demand is expected to re...

    2023

  59. [59]

    [citation]1, 3[/citation] • Newly added intelligent computing power is expected to exceed 100 EFLOPS in

    Development Trends in the Computing Power Market 3.1 Transition Towards Intelligent Computing [citation]1, 3, 13, 14[/citation] 3.1.1 Rapid Growth of Intelligent Computing Power • The CAGR of intelligent computing power, at 46.2%, is significantly higher than the 18.8% CAGR of general-purpose computing power. [citation]1, 3[/citation] • Newly added intell...

  60. [61]

    [citation]11[/citation] • PUE exceeds 2.5 in some regions, indicating low energy efficiency

    Supply and Demand Dynamics 4.1 Structural Imbalances Between Supply and Demand [citation]11[/citation] 4.1.1 Surplus of General-Purpose Computing Power • Some older general-purpose computing centres remain idle, while the average rack deployment rate in internet data centres is only 58%. [citation]11[/citation] • PUE exceeds 2.5 in some regions, indicatin...

    2025

  61. [62]

    Policy Environment and Planning 5.1 National Strategic Planning [citation]6, 10, 19[/citation] 5.1.1 East Data, West Computing Project • The project was launched in 2022 to establish an integrated national computing power network. [citation]5, 6[/citation] • By 2025, China aims to establish urban computing networks with latency of 1 millisecond, regional ...

    2022

  62. [63]

    [citation]17[/citation] • Domestic research and development teams have been forced to slow the development of large language models

    Market Challenges and Risks 6.1 Technical Challenges [citation]5, 17[/citation] 6.1.1 Chip Sanctions and Export Restrictions • The United States has tightened restrictions on the sale of AI chips to China, contributing to a severe shortage of computing power. [citation]17[/citation] • Domestic research and development teams have been forced to slow the de...

  63. [64]

    [citation]2, 4[/citation] • Intelligent computing power capacity is expected to reach 1,460.3 EFLOPS in

    Future Outlook and Opportunities 7.1 Market Size Forecasts [citation]2, 4, 9, 18[/citation] 7.1.1 Near-Term Forecasts, 2025 ⚶2026 • The computing power market is expected to reach RMB 1.05 trillion in 2026. [citation]2, 4[/citation] • Intelligent computing power capacity is expected to reach 1,460.3 EFLOPS in

    2025

  64. [66]

    First, read the prediction

  65. [67]

    The prediction may contain extensive reasoning, intermediate analysis, explanations, or thought processes

  66. [68]

    You must first accurately extract the final predicted answer from the prediction

  67. [69]

    predict_answer

    Then compare the extracted final answer with the ground truth for semantic equivalence. Matching rules: • If both express the same fact, meaning, or conclusion, the result should be considered a match. 45 • Do not question the correctness of the ground truth. • The prediction may contain substantial reflection or reasoning; focus only on the final predict...

  68. [70]

    Locate the sections of the webpage that are directly relevant to the user's information goal

    Content Scanning (Rationale). Locate the sections of the webpage that are directly relevant to the user's information goal

  69. [71]

    Extract all important supporting information without omission

    Evidence Extraction. Extract all important supporting information without omission. Preserve the original context as much as possible rather than paraphrasing. The extracted evidence may span multiple paragraphs when necessary

  70. [72]

    Produce a concise, logically organized summary that prioritizes clarity and explicitly reflects how the extracted information contributes to the user's goal

    Summary Generation. Produce a concise, logically organized summary that prioritizes clarity and explicitly reflects how the extracted information contributes to the user's goal. Constraint Remove distracting or irrelevant content (e.g., advertisements, invalid hyperlinks, navigation elements, and other non-informative text) before performing extraction. O...

  71. [2025]

    [citation]13[/citation] • Lightweight large language models such as DeepSeek are driving rapid expansion in inference applications

    [citation]4[/citation] 3.1.2 Inference Computing as a New Growth Driver • The share of inference computing is expected to rise from 41.3% in 2023 to 73.0% in 2028. [citation]13[/citation] • Lightweight large language models such as DeepSeek are driving rapid expansion in inference applications. [citation]13[/citation] 3.2 Domestic Substitution [citation]5...

    2023

  72. [2026]

    [citation]2, 4[/citation] • Total computing power capacity is expected to exceed 2,500 EFLOPS by 2030

    [citation]22[/citation] 7.1.2 Long-Term Forecasts for 2030 • The computing power market is expected to approach RMB 2 trillion by 2030. [citation]2, 4[/citation] • Total computing power capacity is expected to exceed 2,500 EFLOPS by 2030. [citation]9[/citation] • The direct output of AI-related industries is expected to reach RMB 2 trillion, driving more ...

    2030