Pith. sign in

REVIEW 3 major objections 5 minor 10 cited by

The paper claims that verifiable deep research can be reduced to a hierarchy of constraint-satisfaction subproblems, and that a 3-billion-parameter model trained on 50,000 such synthesized questions outperforms far larger models and several

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 ↗

InfoSeek automatically synthesizes hierarchical constraint-satisfaction QA data from webpages; a 3B agent trained on it reaches 16.5% accuracy on BrowseComp-Plus, outperforming Qwen3-32B and Gemini 2.5 Flash.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Useful open dataset and a credible controlled comparison, but the headline BrowseComp-Plus claim is a 1-point gap without error bars and no decontamination check. the 3 major comments →

arxiv 2509.00375 v1 pith:36J6BT33 submitted 2025-08-30 cs.CL cs.AI

Open Data Synthesis For Deep Research

Ziyi Xia , Kun Luo , Hongjin Qian , Zheng Liu This is my paper
classification cs.CL cs.AI
keywords deep researchhierarchical constraint satisfactiondata synthesisresearch treeagentic searchreinforcement learningopen-source dataset
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 reading

The paper's central claim is that deep research questions with unique, verifiable answers form a distinct problem type—hierarchical constraint satisfaction problems (HCSPs)—that sits above flat constraint combinations and linear multi-hop chains. To train models on this type, the authors build InfoSeek, a fully open pipeline in which two agents grow a Research Tree from Wikipedia and web pages, then blur internal nodes into constraints so that solving the final question requires traversing the whole tree. The pipeline produces over 50K training questions, a curated test set, and 16.5K reasoning trajectories at a reported cost of about $572, with structural depth controlled explicitly. The payoff claimed is empirical: a 3B-parameter model (InfoSeeker) trained on this data with supervised fine-tuning and reinforcement learning scores 16.5% on the 830-problem BrowseComp-Plus benchmark, above 32B open models and several commercial APIs, close to Gemini 2.5 Pro. If the gain is real, it means the scarce ingredient in deep research is not model size but a large supply of structurally constrained, verifiable training problems.

Core claim

On the paper's own terms, the central discovery is that verifiable deep research questions can be generated at scale by converting a tree of entities into a question whose answer requires solving every level of the tree, and that a small model trained on these questions acquires agentic search behavior that transfers to an unseen benchmark. The construction starts with a root entity (the final answer) and recursively attaches child entities and facts drawn from web pages; each internal node is blurred with extra constraints so it becomes a well-defined sub-problem, and an LLM converts the blurred tree into a natural-language question. Quality filters reject questions solvable directly from p

What carries the argument

The load-bearing object is the Research Tree: a tree whose root is the final answer, whose internal vertices are sub-problems, and whose edges are logical dependencies sourced from webpages and Wikipedia. The decisive operation is 'blurring the parent': each internal vertex is given enough child constraints that the vertex is uniquely determined, so the question forces the solver to resolve every layer and prevents shortcuts or memorized answers. Around this, a Planner/Browser dual agent grows the tree, and a quality-assurance step uses a powerful LLM to verify solvability and reject questions answerable from parametric memory. The preserved meta-information—intermediate steps and retrieval

Load-bearing premise

The headline result assumes the 100K webpages used for BrowseComp-Plus evaluation are disjoint from the Wikipedia and web corpora from which InfoSeek training questions were built, so the 3B model cannot have memorized the answers; the paper reports no decontamination check.

What would settle it

Run a decontamination pass: match the root entities and supporting pages of InfoSeek's 50K training trees against the 830 BrowseComp-Plus problems and their fixed 100K corpus, then remove all overlapping samples and retrain. If the 16.5% accuracy is mostly memorization, it should drop toward the NQ+HQA baseline of 3.0%; if InfoSeek teaches a general skill, it should stay well above.

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

If this is right

  • Deep research capability can be synthesized from unstructured text without expensive human annotation, at roughly $572 for the full 50K dataset.
  • Structural complexity of training data can be controlled by vertex count, and measured failure rates rise from 88.1% to 94.1% as trees grow, giving a dial for curriculum design.
  • Because every tree keeps intermediate steps and retrieval labels, training can move beyond outcome rewards to process-level rewards and trajectory-level exploration.
  • Compact 3B models trained this way become practical agents for browsing and search, lowering deployment cost relative to 32B+ systems.

Where Pith is reading between the lines

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

  • The paper does not test this, but the framework's generality could be checked by training on InfoSeek and evaluating on a benchmark whose answer webpages were created after the InfoSeek corpora were frozen; a memorization story predicts a collapse, a general-skill story predicts the gain holds.
  • The HCSP idea likely transfers to other verifiable multi-source tasks—fact-checking against contradictory sources, code archaeology, or literature synthesis—since any task whose answer is unique and decomposable into a tree of sub-answers fits the same synthesis scheme; that extension is the paper's implicit program.
  • An independent human audit of the synthesized questions—checking that they read as natural research questions rather than puzzles only the tree format produces—would show how far the training signal transfers to real user queries.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper introduces InfoSeek, a dual-agent data-synthesis framework that turns Wikipedia and open-web text into Hierarchical Constraint Satisfaction Problems (HCSPs) represented as research trees. The pipeline recursively expands entities, blurs parent nodes with additional constraints, converts trees into natural-language questions, and applies two quality filters. The authors release over 50K QA pairs and 16.5K reasoning trajectories. They train Qwen2.5-3B-Inst with rejection-sampled SFT and GRPO to produce InfoSeeker-3B, and report that on BrowseComp-Plus it reaches 16.5% accuracy, above Gemini 2.5 Flash (15.5%), Sonnet 4 (14.3%), GPT-4.1 (14.6%), and Qwen3-32B (3.5%), while approaching Gemini 2.5 Pro (19.0%). The paper also reports gains on single-hop and multi-hop benchmarks and an ablation showing InfoSeek data outperforms NQ+HQA training data.

Significance. If the empirical claims hold, the paper demonstrates a significant result: a compact, openly released 3B model trained on an openly released synthetic dataset can match or exceed much larger open models and lightweight commercial agents on a challenging deep-research benchmark. The open release of the dataset, trajectories, and synthesis framework is a genuine contribution, and the HCSP formulation is an attractive way to control structural complexity. The strength of the contribution, however, currently rests on a small accuracy margin on a single benchmark and on an unverified assumption that the training corpus and evaluation corpus are disjoint.

major comments (3)
  1. [Sec. 3.1, Sec. A.2, Table 4] The central claim that InfoSeek teaches general deep research assumes that the Wikipedia/web corpus used to synthesize InfoSeek training examples is disjoint from the 100K-page corpus used in BrowseComp-Plus. The manuscript reports no overlap or decontamination analysis. InfoSeek is built from a full Wikipedia dump and webpages (Sec. 3.1), while BrowseComp-Plus uses a fixed 100K web corpus (Sec. A.2); both are drawn from the open web, so overlap is plausible. If the root entities, supporting pages, or answer entities of InfoSeek training trees appear in the BrowseComp-Plus corpus, InfoSeeker-3B could answer from memorized or previously retrieved content, while the commercial APIs and Qwen3-32B have no such exposure. This would invalidate the headline comparison in Table 4 regardless of model size. The authors should report entity-level and page-level overlap statistics between the InfoSe
  2. [Sec. 5.2, Table 4] The claim that InfoSeeker-3B 'surpasses' Gemini 2.5 Flash, Sonnet 4, and GPT-4.1 rests on an absolute gap of about 1.0–2.2 accuracy points on 830 problems. For the 16.5% vs. 15.5% comparison, this is roughly 8 correct answers. No significance testing, confidence intervals, or repeated-run variability is reported. Given the small margin, the difference may be within sampling noise. A McNemar test or paired bootstrap over the 830 questions, and ideally multiple rollout seeds, should be reported before the beyond-chance claim is made. This is load-bearing for the abstract and Sec. 5.2 conclusions.
  3. [Sec. 2.2, Eq. (3), Fig. 2] The formal definition of an HCSP in Eq. (3) does not appear to capture the multi-hop compositionality that the paper uses to distinguish HCSPs from flat CSPs. Eq. (3) defines H(x) as an intersection of constraint-satisfaction sets and recursively defined sub-question solution sets. For the sequential example in Fig. 2b, if the sub-question is 'Who solved the Enigma code?' and the root constraint is 'country whose capital was the birthplace of the Enigma solver,' then the intersection of S(root constraint) and H(sub-question) is empty rather than the intended answer. The composition of intermediate answers into later constraints needs to be represented explicitly; otherwise the formalism reduces every deep research question to a flat set-intersection problem and does not justify the claimed structural distinction. The authors should either fix the recursion (e.g., by making H take and ret
minor comments (5)
  1. [Sec. 3.6, Table 2] The text reports an overall failure rate of 92.7% for Qwen2.5-72B, while Table 2 reports 91.6% for the total. One of these is a typo; please correct and ensure the row-level numbers are consistent with the total.
  2. [Sec. 4.3.1] The GRPO objective uses the denominator |Y| - |K| without defining K or Y. Please define these sets and explain the masking term.
  3. [Abstract, Table 4] The abstract says performance is 'comparable to stronger APIs (e.g., Gemini2.5-Pro).' Table 4 shows Gemini 2.5 Pro at 19.0% vs. 16.5%, and GPT-5 at 55.9%. The claim should be qualified to exclude GPT-5 and to state the exact gap to Gemini 2.5 Pro.
  4. [Sec. A.1 vs. Sec. 4.2] The appendix says the authors 'deliberately retain shortcut cases' among correct trajectories, while Sec. 4.2 says Gemini 2.5 Flash is used to 'check if there exist search or reasoning shortcuts.' Please clarify whether shortcut trajectories are retained or removed, and how this affects the SFT data.
  5. [General] Typos: 'Experienments' in Sec. 5; 'Reserach' in Sec. 4.3.2; 'Condition Satisfactory Problem' and 'Hierarchical Condition Satisfactory Problem' in Fig. 2. Please proofread.

Circularity Check

0 steps flagged

No significant circularity: the headline result is measured on the external BrowseComp-Plus benchmark, while InfoSeek's own synthesis/verification loops are training-data quality gates, not test-set predictions.

full rationale

The paper's central claim (Sec. 5.2, Table 4) is that InfoSeeker-3B, trained on InfoSeek-generated HCSP data, scores 16.5% on BrowseComp-Plus. The evaluation corpus is fixed and external: "we utilize the 100K web page corpus provided by the official release (Chen et al., 2025b), with BM25 serving as the retrieval method" (Sec. A.2). The training objective is defined on InfoSeek's own synthesized questions (Sec. 3) and rewards (Sec. 4.3.2), so no equation makes a training example equal a BrowseComp-Plus test item. The difficulty and verifiability filters (Sec. 3.5) use Qwen2.5-32B and Gemini to remove easy or unsolvable generated questions; these are data-quality gates on the training distribution, not predictions of external benchmark performance. Table 5's comparison of InfoSeek vs NQ+HQA keeps BrowseComp-Plus fixed, so it does not reduce to the training signal. Several citations are to the authors' own work (InForage, BGE-M3, HawkBench, SimpleDeepSearcher), but they are used as baselines, retrievers, or related work, not as load-bearing premises; no uniqueness or correctness claim is imported from them. The one serious validity concern is not by-construction circularity but possible corpus overlap: InfoSeek is built from "webpages and full Wikipedia dump" (Sec. 3.1), while BrowseComp-Plus supplies its own 100K web corpus, and the paper reports no decontamination analysis. If benchmark pages or answer entities leaked into InfoSeek's synthesis corpus, the 16.5% vs 15.5% margin could be inflated by memorization. That is an empirical benchmark-integrity risk requiring overlap measurement, not an exhibited reduction of the paper's derivation to its inputs. Hence no circular step is identified.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical entities. The HCSP formalization and Research Tree are mathematical/data-structure constructs, not invented entities with independent evidence requirements. The main epistemic burden falls on the factual completeness of the web corpus and the faithfulness of LLM-based question generation and verification.

free parameters (2)
  • target tree complexity (vertex count and depth) = 3-7 vertices, with termination at 'desired complexity'
    The Planner's termination condition depends on a chosen complexity objective. The vertex-count distribution (Table 2) shows 3 to 7 vertices, with a small tail at >=7. This is a hand-chosen design parameter that controls question difficulty, not a fitted value.
  • number of blurring constraints per parent node = k claims selected by the Browser
    The Browser selects k claims from a webpage to make the parent entity unique. The choice of k is not specified precisely and varies across examples; it is an ad hoc free parameter in the data generation process.
axioms (4)
  • domain assumption Webpages in the knowledge base contain sufficiently complete factual content to construct valid constraints.
    The whole data pipeline assumes that the Wikipedia dump and webpages provide enough true claims to build unique-answer constraint sets. This is invoked throughout Sec. 3.
  • domain assumption Hyperlinks in webpages reliably indicate logical dependencies between entities.
    Action 3 (extending the tree) treats hyperlinks as evidence of dependency (e.g., 'v was discovered by w'). This is a strong assumption about the semantics of hyperlinks, stated in Sec. 3.3.
  • domain assumption The question-generator LLM (DeepSeek V3 or GPT-4.1) converts research trees into natural-language questions without introducing shortcuts or new information.
    The conversion of trees to questions relies on a powerful LLM being faithful to the tree structure. The paper assumes this faithfulness without measuring it directly, as described in Sec. 3.4.
  • domain assumption Gemini 2.5 Flash verification correctly identifies all underdetermined or overdetermined questions.
    The verifiability filter (Sec. 3.5) trusts Gemini 2.5 Flash to judge whether a question is solvable from the ground-truth pages mixed with distractors. If this judge is imperfect, ambiguous questions may remain in the dataset.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Open Data Synthesis For Deep Research." pith.science (2026). https://pith.science/paper/36J6BT33

@misc{pith2026250900375,
  author       = {Pith},
  title        = {Pith review of: Open Data Synthesis For Deep Research},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/36J6BT33}},
  note         = {Machine review of arXiv:2509.00375}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) are increasingly expected to go beyond simple factual queries toward Deep Research-tasks that require decomposing questions into sub-problems, coordinating multi-step reasoning, and synthesizing evidence from diverse sources. We formalize Deep Research tasks with verifiable answers as Hierarchical Constraint Satisfaction Problems (HCSPs), which are fundamentally different from single-constraint, multi-hop, or flat CSP formulations. However, existing benchmarks (e.g., Natural Questions, HotpotQA) fail to capture this complexity, while recent synthetic datasets often introduce shortcut reasoning, knowledge leakage, or lack sufficient structural depth. To address this gap, we introduce InfoSeek, a scalable framework for synthesizing complex Deep Research tasks. InfoSeek uses a dual-agent system to recursively build a Research Tree from large-scale webpages, blurring intermediate nodes into valid sub-problems, and converting these trees into natural language questions that require traversing the full hierarchy. It also enables rapid scaling, yielding over 50K training examples, a curated test set, and reasoning trajectories generated via reject sampling. Experiments show that models trained on InfoSeek consistently outperform strong baselines. On a challenging benchmark BrowseComp-Plus, 3B LLMs optimized with InfoSeek surpass much larger 32B models and lightweight commercial APIs (e.g., Gemini2.5-Flash), while achieving performance comparable to stronger APIs (e.g., Gemini2.5-Pro). By preserving meta-information such as intermediate steps and retrieval labels, InfoSeek further supports advanced optimization strategies, including compound reward design and trajectory-level exploration. We provide our codes and datasets in \href{https://github.com/VectorSpaceLab/InfoSeek}{this repository}.

Figures

Figures reproduced from arXiv: 2509.00375 by Hongjin Qian, Kun Luo, Zheng Liu, Ziyi Xia.

Figure 1
Figure 1. Figure 1: Performance comparison on the BrowseComp-Plus benchmark. InfoSeeker-3B, a com [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of Constraint Satisfaction Problems (CSP), Multi-hop Problems (MHP), and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Framework of InfoSeeker, which decomposes target tasks, conducts step-by-step investi [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Statistics for SFT trajectory data. A.4 ANNOUNCEMENT The code and data accompanying this work are released under the Apache License, Version 2.0. This permits use, modification, and distribution for research and commercial purposes, provided that proper attribution is given and the terms of the license are followed. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Research tree structure of Case One in InfoSeek. [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Research tree structure of Case Two in InfoSeek. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 10 Pith papers

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

  1. FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents

    cs.CL 2026-06 unverdicted novelty 7.0

    FORT synthesizes shortcut-resistant search tasks by controlling four identified shortcut risks across entity selection, graph construction, question formulation, and refinement, producing training data that yields age...

  2. HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents

    cs.LG 2026-05 unverdicted novelty 7.0

    HyperEyes uses a dual-grained RL framework with parallel tool actions and efficiency rewards to achieve 9.9% higher accuracy and 5.3x fewer tool calls than prior open-source multimodal agents.

  3. Scaling the Scaling Logic: Agentic Meta-Synthesis of Logic Reasoning

    cs.AI 2026-01 unverdicted novelty 7.0

    SSLogic uses LLM agents in a closed Generate-Validate-Refine loop to evolve 953 logic task families from 400 seeds, producing data that yields benchmark gains of +5.2 on SynLogic, +3.0 on AIME25, and +5.5 on BBH.

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

    cs.AI 2026-08 conditional novelty 6.0

    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.

  5. HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents

    cs.LG 2026-05 unverdicted novelty 6.0

    HyperEyes presents a parallel multimodal search agent using dual-grained efficiency-aware RL with a new TRACE reward and IMEB benchmark, claiming 9.9% higher accuracy and 5.3x fewer tool calls than prior open-source agents.

  6. POINTS-Seeker: An Open Recipe for Multimodal Search Agents with Visual Memory Management

    cs.CV 2026-04 conditional novelty 6.0

    An 8B multimodal search model trained with Agentic Seeding and V-Fold visual-history compression reaches state-of-the-art scores on six web-search VQA benchmarks.

  7. POINTS-Seeker: An Open Recipe for Multimodal Search Agents with Visual Memory Management

    cs.CV 2026-04 unverdicted novelty 6.0

    POINTS-Seeker-8B is an 8B multimodal model trained from scratch for agentic search that uses seeding and visual-space history folding to outperform prior models on six visual reasoning benchmarks.

  8. Learning to Retrieve from Agent Trajectories

    cs.IR 2026-03 conditional novelty 6.0

    Retrievers trained on agent trajectories via the LRAT framework improve evidence recall, task success, and efficiency in agentic search benchmarks.

  9. Lost in the Maze: Overcoming Context Limitations in Long-Horizon Agentic Search

    cs.CL 2025-10 conditional novelty 6.0

    SLIM separates search and browse tools and summarizes trajectories every 50 turns, beating several open-source agentic search systems on BrowseComp and HLE with fewer tool calls and lower cost.

  10. Agentic Environment Engineering for Large Language Models: A Survey of Environment Modeling, Synthesis, Evaluation, and Application

    cs.CL 2026-06 unverdicted novelty 5.0

    This survey categorizes agentic environments for LLMs by eight attributes and domains, introduces symbolic and neural synthesis paradigms with evaluation, and outlines four agent evolution pathways plus three environm...

Reference graph

Works this paper leans on

34 extracted references · 4 canonical work pages · cited by 8 Pith papers

  1. [4]

    Next, we conduct the second stage of reinforcement learning

    Using the same hy- perparameters as Round 1, we obtain InfoSeeker-3B-SFT. Next, we conduct the second stage of reinforcement learning. From the original 55K data pool, we select 17K harder samples (15K In- foSeek, 2K NQ & HQA). Before training, the model generates preliminary answers, and we keep only 14K samples it fails on. Using GRPO without KL loss fo...

    work page 2023

  2. [6]

    Cognitive kernel-pro: A framework for deep research agents and agent foundation models training

    13 Technical report Tianqing Fang, Zhisong Zhang, Xiaoyang Wang, Rui Wang, Can Qin, Yuxuan Wan, Jun-Yu Ma, Ce Zhang, Jiaqi Chen, Xiyun Li, et al. Cognitive kernel-pro: A framework for deep research agents and agent foundation models training. arXiv preprint arXiv:2508.00414,

    Pith/arXiv arXiv

  3. [7]

    Qwen Group

    URL https://arxiv.org/ abs/2507.06261. Qwen Group. Qwen2.5 technical report,

    Pith/arXiv arXiv

  4. [8]

    URL https://arxiv.org/abs/2412. 15115. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948,

    Pith/arXiv arXiv

  5. [9]

    Deep researcher with test-time diffusion

    Rujun Han, Yanfei Chen, Zoey CuiZhu, Lesly Miculicich, Guan Sun, Yuanjun Bi, Weiming Wen, Hui Wan, Chunfeng Wen, Sol`ene Maˆıtre, et al. Deep researcher with test-time diffusion. arXiv preprint arXiv:2507.16075,

    Pith/arXiv arXiv

  6. [10]

    Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060 ,

    Pith/arXiv arXiv 2011

  7. [12]

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

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Dong Wang, Hamed Zamani, and Jiawei Han. Search- r1: Training llms to reason and leverage search engines with reinforcement learning. CoRR, abs/2503.09516,

    Pith/arXiv arXiv

  8. [13]

    URL https://doi.org/10

    doi: 10.48550/ARXIV .2503.09516. URL https://doi.org/10. 48550/arXiv.2503.09516. Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551,

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.09516

  9. [15]

    14 Technical report Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al

    Accessed: 2025-08-24. 14 Technical report Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to fol- low instructions with human feedback. Advances in neural information processing systems , 35: 27730–27744,

    work page 2025

  10. [16]

    Measuring and narrowing the compositionality gap in language models

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

    Pith/arXiv arXiv

  11. [18]

    Hawkbench: Investigating resilience of rag methods on stratified information-seeking tasks

    Hongjin Qian, Zheng Liu, Chao Gao, Yankai Wang, Defu Lian, and Zhicheng Dou. Hawkbench: Investigating resilience of rag methods on stratified information-seeking tasks. arXiv preprint arXiv:2502.13465,

    work page arXiv

  12. [19]

    Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution

    Jiahao Qiu, Xuan Qi, Tongcheng Zhang, Xinzhe Juan, Jiacheng Guo, Yifu Lu, Yimin Wang, Zixin Yao, Qihan Ren, Xun Jiang, et al. Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution. arXiv preprint arXiv:2505.20286,

    Pith/arXiv arXiv

  13. [23]

    Pangu deepdiver: Adaptive search intensity scaling via open-web reinforcement learning

    Wenxuan Shi, Haochen Tan, Chuqiao Kuang, Xiaoguang Li, Xiaozhe Ren, Chen Zhang, Hanting Chen, Yasheng Wang, Lifeng Shang, Fisher Yu, et al. Pangu deepdiver: Adaptive search intensity scaling via open-web reinforcement learning. arXiv preprint arXiv:2505.24332, 2025a. Yaorui Shi, Sihang Li, Chang Wu, Zhiyuan Liu, Junfeng Fang, Hengxing Cai, An Zhang, and X...

    arXiv

  14. [24]

    Zerosearch: Incentivize the search capability of llms without searching

    Hao Sun, Zile Qiao, Jiayan Guo, Xuanbo Fan, Yingyan Hou, Yong Jiang, Pengjun Xie, Yan Zhang, Fei Huang, and Jingren Zhou. Zerosearch: Incentivize the search capability of llms without searching. arXiv preprint arXiv:2505.04588, 2025a. Shuang Sun, Huatong Song, Yuhao Wang, Ruiyang Ren, Jinhao Jiang, Junjie Zhang, Fei Bai, Jia Deng, Wayne Xin Zhao, Zheng Li...

    Pith/arXiv arXiv

  15. [25]

    Qwen2 technical report

    Qwen Team. Qwen2 technical report. arXiv preprint arXiv:2407.10671,

    Pith/arXiv arXiv

  16. [26]

    Interleaving re- trieval with chain-of-thought reasoning for knowledge-intensive multi-step questions

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Interleaving re- trieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. arXiv preprint arXiv:2212.10509, 2022a. Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Musique: Multihop questions via single-hop question compositio...

    Pith/arXiv arXiv

  17. [27]

    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,

    Pith/arXiv arXiv

  18. [28]

    Webdancer: Towards autonomous information seeking agency, 2025a

    Jialong Wu, Baixuan Li, Runnan Fang, Wenbiao Yin, Liwen Zhang, Zhengwei Tao, Dingchu Zhang, Zekun Xi, Gang Fu, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. Webdancer: Towards autonomous information seeking agency, 2025a. URL https://arxiv.org/abs/2505. 22648. Jialong Wu, Wenbiao Yin, Yong Jiang, Zhenglin Wang, Zekun Xi, Runnan Fang, Linhai Zhang,...

    Pith/arXiv arXiv

  19. [29]

    Qwen3 technical report

    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,

    Pith/arXiv arXiv

  20. [30]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering

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

    Pith/arXiv arXiv

  21. [31]

    Agentic information retrieval

    Weinan Zhang, Junwei Liao, Ning Li, Kounianhua Du, and Jianghao Lin. Agentic information retrieval. arXiv preprint arXiv:2410.09713,

    Pith/arXiv arXiv

  22. [32]

    Agentorches- tra: A hierarchical multi-agent framework for general-purpose task solving

    Wentao Zhang, Ce Cui, Yilei Zhao, Rui Hu, Yang Liu, Yahui Zhou, and Bo An. Agentorches- tra: A hierarchical multi-agent framework for general-purpose task solving. arXiv preprint arXiv:2506.12508,

    Pith/arXiv arXiv

  23. [33]

    Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu

    URL https://arxiv.org/abs/2409.14924. 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 environ- ments. arXiv preprint arXiv:2504.03160,

    Pith/arXiv arXiv

  24. [34]

    shortcut

    executing research workflows proposed in Sec 4.1, which are then used for SFT of Qwen2.5- 3B-Inst (Group, 2025). Concretely, we utilize 50K InfoSeek samples (For training advanced deep research capability) and 5K NQ & HQA samples (For preserving general agentic search capability), each rolled out twice. After filtering incorrect executions, we obtain 24K ...

    work page 2025

  25. [1976]

    RQ- RAG: learning to refine queries for retrieval augmented generation.CoRR, abs/2404.00610,

    Chi-Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yike Guo, and Jie Fu. RQ- RAG: learning to refine queries for retrieval augmented generation.CoRR, abs/2404.00610,

    Pith/arXiv arXiv

  26. [2009]

    High- dimensional continuous control using generalized advantage estimation

    John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High- dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438,

    Pith/arXiv arXiv

  27. [2015]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347,

    Pith/arXiv arXiv

  28. [2017]

    Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models

    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 mathemati- cal reasoning in open language models. arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  29. [2019]

    Websailor: Navigating super-human reasoning for web agent

    Kuan Li, Zhongwang Zhang, Huifeng Yin, Liwen Zhang, Litu Ou, Jialong Wu, Wenbiao Yin, Baix- uan Li, Zhengwei Tao, Xinyu Wang, et al. Websailor: Navigating super-human reasoning for web agent. arXiv preprint arXiv:2507.02592, 2025a. Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Search-o1: Agentic...

    Pith/arXiv arXiv

  30. [2020]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720,

    Pith/arXiv arXiv

  31. [2022]

    Scent of knowledge: Optimizing search-enhanced reasoning with information foraging

    Hongjin Qian and Zheng Liu. Scent of knowledge: Optimizing search-enhanced reasoning with information foraging. arXiv preprint arXiv:2505.09316,

    arXiv

  32. [2023]

    Learning to reason with search for llms via reinforce- ment learning

    Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, Fan Yang, et al. Learning to reason with search for llms via reinforce- ment learning. arXiv preprint arXiv:2503.19470, 2025a. Zijian Chen, Xueguang Ma, Shengyao Zhuang, Ping Nie, Kai Zou, Andrew Liu, Joshua Green, Kshama Patel, Ruoxi Meng, M...

    Pith/arXiv arXiv

  33. [2024]

    URL https://doi.org/10.48550/arXiv.2404

    doi: 10.48550/ARXIV .2404.00610. URL https://doi.org/10.48550/arXiv.2404. 00610. Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge dis- tillation,

    Pith/arXiv arXiv doi:10.48550/arxiv.2404.00610

  34. [2025]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi

    Ac- cessed: 2025-08-24. Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511,

    Pith/arXiv arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.