Pith. sign in

REVIEW 3 major objections 6 minor 89 references

A single data pipeline that covers text-only, image-to-text, and text-to-image browsing, then keeps only high-exploration questions for RL, trains a 35B agent to 54.4% average on five multimodal BrowseComp benchmarks.

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 · grok-4.5

2026-07-14 10:49 UTC pith:YD6WYOK6

load-bearing objection Solid open-source systems paper that actually fills the text-to-image BrowseComp gap and posts real gains; the abstract’s closed-source comparison is selective and the exploration-degree story is thinner than the pattern-coverage story. the 3 major comments →

arxiv 2607.10557 v1 pith:YD6WYOK6 submitted 2026-07-12 cs.CL

UNIBROWSE: A Data-to-Agent Framework for Multimodal BrowseComp

Xiyu Wei , Qingwei Zong , Zhuocheng Yu , Sujian Li This is my paper
classification cs.CL
keywords multimodal BrowseCompweb browsing agentstext-to-image patternexploration degreeknowledge-graph data synthesistool-use trajectoriesreinforcement learningsupervised fine-tuning
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.

Multimodal web browsing is not one skill: sometimes the answer is pure text composition, sometimes an image clue must be resolved first, and sometimes textual search only tells you which visual detail still needs checking. Prior training pipelines mostly built the first two patterns and left the third undertrained, and they fed reinforcement learning whatever synthetic problems they generated, including ones with almost no useful search diversity. UNIBROWSE builds questions from knowledge-graph skeletons expanded with live web attributes, materializes all three information-flow patterns from the same skeleton when feasible, and scores each question by how much successful base-agent rollouts vary in browsing effort. Keeping the high-variance questions for RL, and cold-start trajectories for supervised fine-tuning, produces a 35B agent that gains 10.5 points over its base model and exceeds several closed-source agent workflows on the reported suite. A sympathetic reader cares because the claim is that pattern coverage plus exploration filtering, not just larger models or more unfiltered rollouts, is what turns synthetic browse data into usable long-horizon multimodal agents.

Core claim

The paper claims that jointly generating training data for all three real multimodal browsing patterns (text-only, image-to-text, and text-to-image), grounding graph skeletons with verified live-web attributes, and selecting reinforcement-learning instances by exploration degree yields a 35B-scale agent that reaches 54.4 average accuracy across five multimodal BrowseComp benchmarks—10.5 points above its Qwen3.5-35B-A3B base and above several closed-source agent workflows such as GPT-5, Gemini-2.5 Pro, and Gemini-2.5 Flash.

What carries the argument

Exploration degree E(q): the variance, across successful base-agent rollouts, of browsing-step counts normalized by the question’s fixed set of key evidence steps. It is used only to keep high-signal RL questions (top 30% in the final agent), not as a reward. The complementary mechanism is unified pattern materialization: the same web-augmented layered evidence subgraph can attach an entry image, a final visual-attribute image, both, or neither.

Load-bearing premise

That variance in how long successful rollouts take is a reliable filter for questions that actually improve the policy under group-relative RL, rather than merely dropping easy, hard, or noisy items.

What would settle it

Retrain the same SFT checkpoint with an equal-sized RL set sampled at random (or by difficulty proxies other than E(q)) from the identical validated pool; if the top-30% E(q) set no longer wins on MM-BrowseComp and BrowseComp-VL under the paper’s reported setup, the exploration-selection claim fails.

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

If this is right

  • Multimodal browse agents trained without an explicit text-to-image construction path will keep lagging on mixed-pattern benchmarks such as MM-BrowseComp.
  • Graph-only or random-walk-only data generators can be replaced by hybrid skeletons that add live-web attributes without abandoning controllable multi-hop structure.
  • RL budgets for long-horizon browsing can be spent on a smaller high-E(q) pool instead of unfiltered synthetic QA, improving accuracy per training step.
  • Cold-start tool trajectories plus exploration-aware GRPO on a mid-size open base can close a large fraction of the gap to proprietary browse workflows on public image-and-web tools.
  • Future data pipelines for browse agents should report pattern mix and exploration-selection ablations, not only raw trajectory counts.

Where Pith is reading between the lines

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

  • The same E(q)-style variance filter may transfer to non-visual multi-hop tool agents (code, maps, APIs) wherever successful trajectories have measurable length diversity.
  • If web content drifts, fixed offline E(q) scores will stale; periodic re-rollout of the filter set is a natural maintenance cost the paper leaves open.
  • Hybrid questions that combine entry-image and final-image nodes may be the highest-leverage subset for teaching modality switching, even when the three patterns are balanced 1:1:1.
  • Proprietary workflows that still lead on average may be winning mainly on tool stack and foundation scale rather than on pattern coverage alone, given how much the open 35B gains from data design.

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 / 6 minor

Summary. UNIBROWSE presents a data-to-agent framework for multimodal BrowseComp. Its pipeline starts from textual knowledge-graph skeletons, expands them with live web attributes, materializes three information-flow patterns (text-only, image-to-text, and the previously under-addressed text-to-image), and filters RL instances by an exploration-degree metric E(q) defined as the variance of normalized browsing-step counts over successful base-agent rollouts. From this pipeline the authors produce ~8K cold-start tool-use trajectories and ~10K exploration-rich QA pairs, then train a Qwen3.5-35B-A3B agent with SFT followed by GRPO. The resulting agent reaches 54.4 average accuracy across five public benchmarks (BC-VL, MM-BC, MMSearch+, MMSearch, FVQA), a +10.5 gain over the base model, with ablations in Tables 3–4 supporting pattern coverage and the top-30% E(q) selection rule.

Significance. Multimodal open-web browsing with compositional, multi-hop evidence is a timely and difficult setting; covering text-to-image as a first-class construction pattern is a genuine gap-filling contribution relative to WebWatcher, Vision-DeepResearch, and Skywork-R1V4. The hybrid KG-plus-live-web construction and the explicit exploration-aware RL filter are practically useful design choices, and the SFT→RL gains plus pattern ablations give a coherent empirical story for open-source agents. Strengths include a clearly specified tool protocol, transparent SFT vs SFT+RL reporting, and public-benchmark evaluation rather than in-distribution test splits. If the comparison framing and the interpretation of E(q) are tightened, the work is a solid systems contribution for multimodal agent post-training.

major comments (3)
  1. Abstract and Figure 1 present the 54.4 average as surpassing closed-source agent workflows (GPT-5 42.9, Gemini-2.5 Pro 44.8, Gemini-2.5 Flash 41.3) while omitting stronger proprietary workflows that Table 2 fully reports—Gemini-3.1 Pro (63.2) and Doubao Seed 2.0 Pro (62.5). The body (§5.2) correctly notes the remaining gap to 63.2, but the abstract/Figure 1 comparison set is selective and load-bearing for the headline claim. Please reframe the abstract and Figure 1 to state open-source SOTA (or list the full comparable set), and make the Avg. column’s inclusion rule explicit whenever models lack MM-BC scores.
  2. §5.1 evaluates MM-BC only on the non-video, non-map subset so that all systems share the same image+web tools. MM-BC is the sole benchmark that exercises all three information-flow patterns—the paper’s central novelty—so subset size, pattern composition (text-only / I2T / T2I counts), and any score shift relative to the full set are load-bearing. Please report the subset cardinality and pattern mix, and discuss whether the 28.2 result can be read as evidence for full-pattern generality.
  3. §3.3 and Eq. (3) define exploration degree E(q) as Var[ℓ(τb, π)] over successful rollouts and treat high variance as richer GRPO signal. Table 4 peaks at top-30% but is non-monotonic (top-10% collapses), and the manuscript does not show that E(q) is not merely a proxy for difficulty, answer rarity, or base-agent instability. Because exploration-aware selection is a named contribution and the sole filter for the expensive RL stage, please add (i) correlations of E(q) with success rate / |π| / trajectory entropy and (ii) a control that matches difficulty while randomizing E(q), or else soften the causal claim that variance of browsing length is what improves policy gradients.
minor comments (6)
  1. Abstract: typo “serveral” → “several”; also “UNIBROWSE” spacing is inconsistent (UNIBROWSE / UNI-BROWSE / UniBrowse).
  2. §5.1: “full training details provided in Appendix??” is an unresolved placeholder.
  3. Appendix B.2 duplicates the Reward Design and Policy Optimization paragraphs almost verbatim; collapse to a single copy.
  4. Table 2 / metrics: LLM-as-a-judge accuracy is used without reporting judge model, prompt, or agreement with human labels; a short reliability note would strengthen the numbers.
  5. Free parameters (γ, Nmin, λ_ans/λ_fmt, B rollouts for E(q)) are stated in the appendices but not sensitivity-tested beyond the E(q) quantile; even a brief note would help reproducibility.
  6. Figure 1 y-axis scales differ across panels, which makes visual comparison of the Avg. bar harder; a shared scale or explicit callout would help.

Circularity Check

0 steps flagged

No circularity: empirical data-to-agent pipeline evaluated on external public benchmarks; exploration degree is a selection filter only.

full rationale

UNIBROWSE is a systems paper whose central claims are (i) a constructive data pipeline that materializes three information-flow patterns from KG skeletons plus live web attributes and (ii) an empirical SOTA accuracy of 54.4 on five external multimodal BrowseComp benchmarks after SFT+GRPO. Neither claim reduces by construction to its inputs. The exploration-degree statistic E(q) = Var[ℓ(τb, π)] is used solely to rank and retain the top-30 % of QA pairs for the RL stage (Section 3.3, Table 4); it never enters the reward R = λ_ans R_ans + λ_fmt R_fmt, so the reported accuracy is not forced by the filter. Teacher trajectories are generated by an external model (Kimi K2.5), quality checks are independent sufficiency/uniqueness tests, and all five evaluation suites (BC-VL, MM-BC subset, MMSearch+, MMSearch, FVQA) are public and external. No self-definitional equations, fitted-parameter-as-prediction, load-bearing self-citation of uniqueness theorems, or renaming of known results appear. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 1 invented entities

The central performance claim rests on a handful of hand-chosen thresholds (top-30% E(q), support count γ=3, Nmin=5 pages) and on domain assumptions that synthetic KG-plus-web questions transfer to real benchmarks and that variance of successful rollout lengths is a valid proxy for RL signal. No new physical entities are postulated; the only invented construct is the exploration-degree statistic itself.

free parameters (4)
  • exploration-degree retention quantile = top 30%
    Top-30% threshold chosen after ablation (Table 4); final accuracy is sensitive to this cut-off.
  • web-attribute support threshold γ = 3
    Attribute kept only if corroborated by ≥3 independent pages (Appendix C); controls noise vs. coverage.
  • minimum usable pages Nmin = 5
    Entity must have ≥5 usable pages before attribute extraction (Appendix C).
  • reward weights λ_ans, λ_fmt
    Composite GRPO reward coefficients; exact values not numerically reported but required for the RL stage.
axioms (3)
  • domain assumption Layered knowledge-graph subgraphs plus live web attributes constitute a faithful proxy for real multimodal BrowseComp evidence.
    Invoked throughout Section 3; without transfer from synthetic construction to public benchmarks the performance claim collapses.
  • ad hoc to paper Variance of normalized browsing-step counts across successful rollouts measures the richness of the learning signal for GRPO.
    Definition of E(q) in Section 3.3; the paper treats high variance as desirable without independent theoretical justification.
  • domain assumption An MLLM judge can reliably decide entity identifiability, extract non-textual visual attributes, and enforce modality necessity.
    Used to construct Ut2i(a) (Appendix B); errors here would inject invalid text-to-image instances.
invented entities (1)
  • exploration degree E(q) no independent evidence
    purpose: Scalar filter that ranks QA pairs by diversity of successful browsing trajectories so that only high-signal instances enter RL.
    Defined in Eq. (3) as Var[ℓ(τb,π)]; no prior literature cited for this exact statistic; independent_evidence is false because it is validated only by the same downstream accuracy numbers it is meant to improve.

pith-pipeline@v1.1.0-grok45 · 21835 in / 3077 out tokens · 44028 ms · 2026-07-14T10:49:47.413017+00:00 · methodology

0 comments
read the original abstract

Multimodal BrowseComp tasks require agents to combine perception, tool use, and long-horizon reasoning over dynamic web content, challenging their ability to handle compositional structure, open-world uncertainty, and multimodal integration across extended interactions. Crucially, real-world multimodal browsing involves three distinct information-flow patterns: text-only, image-to-text, and text-to-image, yet existing data construction methods cover only the text-only and image-to-text patterns, leaving text-to-image largely unaddressed and limiting agent generality and robustness. We introduce UNIBROWSE, a unified data pipeline that for the first time simultaneously generates training data covering all three patterns, augments curated knowledge graphs with live web retrieval for improved fidelity, and introduces a novel metric of exploration degree to filter low-signal instances for efficient reinforcement learning. Through this pipeline, we produce high-quality cold-start tool-use trajectories and exploration-rich QA pairs, and train a 35B-scale agent via supervised fine-tuning and exploration-aware RL.The resulting UNIBROWSE agent achieves state-of-the-art performance on multimodal BrowseComp benchmarks, attaining an average accuracy of 54.4 across five diverse benchmarks -- an improvement of 10.5 points over its base model Qwen3.5-35B-A3B -- and surpassing serveral closed-source agent workflows such as GPT-5 (42.9), Gemini-2.5 Pro (44.8), and Gemini-2.5 Flash (41.3).

Figures

Figures reproduced from arXiv: 2607.10557 by Qingwei Zong, Sujian Li, Xiyu Wei, Zhuocheng Yu.

Figure 1
Figure 1. Figure 1: Overall performance of UNIBROWSE Agent compared to other models across five benchmarks. from restricted web diversity, while random-walk methods expose data generation to massive noise and false evidence. Third, prior works apply learn￾ing directly on the generated data without a fil￾tering mechanism tailored for the long-horizon, multi-step nature of BrowseComp reasoning. Con￾sequently, low-quality trajec… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the UNIBROWSE framework. (a) The data construction pipeline starts from textual knowledge-graph skeletons, augments them with live web evidence, expands them into text-only, image-to-text, and text-to-image patterns, and applies exploration-aware selection to RL QA pairs. (b) The UNIBROWSE multimodal browsing trajectory shows how the agent resolves an image clue, follows textual web evidence, a… view at source ↗
Figure 3
Figure 3. Figure 3: A case study of UNIBROWSE on a multimodal browsing question [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Prompt used to extract candidate attributes from a single webpage summary. [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Prompt used to aggregate and verify repeated attributes across webpages. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

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

89 extracted references · 1 canonical work pages

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  2. [2]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  3. [3]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  4. [4]

    arXiv preprint arXiv:2404.06654 , year=

    RULER: What's the Real Context Size of Your Long-Context Language Models? , author=. arXiv preprint arXiv:2404.06654 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2406.10149 , year=

    BABILong: Testing the Limits of LLMs with Long Context Reasoning-in-a-Haystack , author=. arXiv preprint arXiv:2406.10149 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2412.15204 , year=

    LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks , author=. arXiv preprint arXiv:2412.15204 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    Pith/arXiv arXiv

  8. [8]

    5 Technical Report , author=

    Qwen2. 5 Technical Report , author=. arXiv preprint arXiv:2412.15115 , year=

    Pith/arXiv arXiv

  9. [9]

    Claude 3.5 Sonnet , howpublished =

    Claude, Team , year=. Claude 3.5 Sonnet , howpublished =

  10. [10]

    QwQ , howpublished =

    QwQ, Team , year=. QwQ , howpublished =

  11. [11]

    Hello GPT-4o , howpublished =

    GPT4o, Team , year=. Hello GPT-4o , howpublished =

  12. [12]

    arXiv preprint arXiv:2409.12183 , year=

    To cot or not to cot? chain-of-thought helps mainly on math and symbolic reasoning , author=. arXiv preprint arXiv:2409.12183 , year=

    Pith/arXiv arXiv

  13. [13]

    Transactions of the Association for Computational Linguistics , volume=

    MuSiQue: Multihop Questions via Single-hop Question Composition , author=. Transactions of the Association for Computational Linguistics , volume=. 2022 , publisher=

    2022

  14. [14]

    arXiv preprint arXiv:2407.10671 , year=

    Qwen2 Technical Report , author=. arXiv preprint arXiv:2407.10671 , year=

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:1809.09600 , year=

    HotpotQA: A dataset for diverse, explainable multi-hop question answering , author=. arXiv preprint arXiv:1809.09600 , year=

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2011.01060 , year=

    Constructing a multi-hop QA dataset for comprehensive evaluation of reasoning steps , author=. arXiv preprint arXiv:2011.01060 , year=

    Pith/arXiv arXiv 2011

  17. [17]

    arXiv preprint arXiv:2308.14508 , year=

    Longbench: A bilingual, multitask benchmark for long context understanding , author=. arXiv preprint arXiv:2308.14508 , year=

    Pith/arXiv arXiv

  18. [18]

    arXiv preprint arXiv:2410.02660 , year=

    How to Train Long-Context Language Models (Effectively) , author=. arXiv preprint arXiv:2410.02660 , year=

    arXiv

  19. [19]

    Forty-first International Conference on Machine Learning , year=

    Data Engineering for Scaling Language Models to 128K Context , author=. Forty-first International Conference on Machine Learning , year=

  20. [20]

    Long Context is Not Long at All: A Prospector of Long-Dependency Data for Large Language Models

    Chen, Longze and Liu, Ziqiang and He, Wanwei and Zheng, Yinhe and Sun, Hao and Li, Yunshui and Luo, Run and Yang, Min. Long Context is Not Long at All: A Prospector of Long-Dependency Data for Large Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.a...

    work page doi:10.18653/v1/2024.acl-long.447 2024

  21. [21]

    ArXiv , year=

    Extending Context Window of Large Language Models via Positional Interpolation , author=. ArXiv , year=

  22. [22]

    The Twelfth International Conference on Learning Representations , year=

    PoSE: Efficient Context Window Extension of LLMs via Positional Skip-wise Training , author=. The Twelfth International Conference on Learning Representations , year=

  23. [23]

    The Twelfth International Conference on Learning Representations , year=

    YaRN: Efficient Context Window Extension of Large Language Models , author=. The Twelfth International Conference on Learning Representations , year=

  24. [24]

    Effective Long-Context Scaling of Foundation Models

    Xiong, Wenhan and Liu, Jingyu and Molybog, Igor and Zhang, Hejia and Bhargava, Prajjwal and Hou, Rui and Martin, Louis and Rungta, Rashi and Sankararaman, Karthik Abinav and Oguz, Barlas and Khabsa, Madian and Fang, Han and Mehdad, Yashar and Narang, Sharan and Malik, Kshitiz and Fan, Angela and Bhosale, Shruti and Edunov, Sergey and Lewis, Mike and Wang,...

    doi:10.18653/v1/2024.naacl-long.260 2024

  25. [25]

    Forty-first International Conference on Machine Learning , year=

    LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens , author=. Forty-first International Conference on Machine Learning , year=

  26. [26]

    Selecting Influential Samples for Long Context Alignment via Homologous Models’ Guidance and Contextual Awareness Measurement , author=

  27. [27]

    arXiv preprint arXiv:2410.18745 , year=

    Why Does the Effective Context Length of LLMs Fall Short? , author=. arXiv preprint arXiv:2410.18745 , year=

    Pith/arXiv arXiv

  28. [28]

    Forty-first International Conference on Machine Learning , year=

    LLM Maybe LongLM: SelfExtend LLM Context Window Without Tuning , author=. Forty-first International Conference on Machine Learning , year=

  29. [29]

    arXiv preprint arXiv:2412.18860 , year=

    Bootstrap Your Own Context Length , author=. arXiv preprint arXiv:2412.18860 , year=

    Pith/arXiv arXiv

  30. [30]

    ArXiv , year=

    What are the Essential Factors in Crafting Effective Long Context Multi-Hop Instruction Datasets? Insights and Best Practices , author=. ArXiv , year=

  31. [31]

    L ong A lign: A Recipe for Long Context Alignment of Large Language Models

    Bai, Yushi and Lv, Xin and Zhang, Jiajie and He, Yuze and Qi, Ji and Hou, Lei and Tang, Jie and Dong, Yuxiao and Li, Juanzi. L ong A lign: A Recipe for Long Context Alignment of Large Language Models. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.74

    doi:10.18653/v1/2024.findings-emnlp.74 2024

  32. [32]

    arXiv preprint arXiv:2405.03939 , year=

    Long context alignment with short instructions and synthesized positions , author=. arXiv preprint arXiv:2405.03939 , year=

    Pith/arXiv arXiv

  33. [33]

    L ong E mbed: Extending Embedding Models for Long Context Retrieval

    Zhu, Dawei and Wang, Liang and Yang, Nan and Song, Yifan and Wu, Wenhao and Wei, Furu and Li, Sujian. L ong E mbed: Extending Embedding Models for Long Context Retrieval. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp-main.47

    doi:10.18653/v1/2024.emnlp-main.47 2024

  34. [34]

    arXiv preprint arXiv:2404.02060 , year=

    Long-context llms struggle with long in-context learning , author=. arXiv preprint arXiv:2404.02060 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2402.14848 , year=

    Same task, more tokens: the impact of input length on the reasoning performance of large language models , author=. arXiv preprint arXiv:2402.14848 , year=

    Pith/arXiv arXiv

  36. [36]

    GitHub repository , howpublished =

    Kamradt, Greg , title =. GitHub repository , howpublished =. 2023 , publisher =

    2023

  37. [37]

    Eric Zelikman and Yuhuai Wu and Jesse Mu and Noah Goodman , booktitle=

  38. [38]

    Arian Hosseini and Xingdi Yuan and Nikolay Malkin and Aaron Courville and Alessandro Sordoni and Rishabh Agarwal , booktitle=. V-

  39. [39]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Iterative Reasoning Preference Optimization , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  40. [40]

    arXiv preprint arXiv:2406.03816 , year=

    ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search , author=. arXiv preprint arXiv:2406.03816 , year=

    Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2308.08998 , year=

    Reinforced self-training (rest) for language modeling , author=. arXiv preprint arXiv:2308.08998 , year=

    Pith/arXiv arXiv

  42. [42]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Self-Training with Direct Preference Optimization Improves Chain-of-Thought Reasoning , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  43. [43]

    arXiv preprint arXiv:2312.06585 , year=

    Beyond human data: Scaling self-training for problem-solving with language models , author=. arXiv preprint arXiv:2312.06585 , year=

    Pith/arXiv arXiv

  44. [44]

    arXiv preprint arXiv:2411.08147 , year=

    Large Language Models Can Self-Improve in Long-context Reasoning , author=. arXiv preprint arXiv:2411.08147 , year=

    Pith/arXiv arXiv

  45. [45]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  46. [46]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , address=

    LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , address=. 2024 , url=

    2024

  47. [47]

    Advances in neural information processing systems , volume=

    Language models are few-shot learners , author=. Advances in neural information processing systems , volume=

  48. [48]

    The N arrative QA Reading Comprehension Challenge

    Ko c isk \'y , Tom \'a s and Schwarz, Jonathan and Blunsom, Phil and Dyer, Chris and Hermann, Karl Moritz and Melis, G \'a bor and Grefenstette, Edward. The N arrative QA Reading Comprehension Challenge. Transactions of the Association for Computational Linguistics. 2018. doi:10.1162/tacl_a_00023

    doi:10.1162/tacl_a_00023 2018

  49. [49]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    QMSum: A New Benchmark for Query-based Multi-domain Meeting Summarization , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  50. [50]

    arXiv preprint arXiv:2406.11612 , year=

    Long Code Arena: a Set of Benchmarks for Long-Context Code Models , author=. arXiv preprint arXiv:2406.11612 , year=

    Pith/arXiv arXiv

  51. [51]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  52. [52]

    arXiv preprint arXiv:2209.07686 , year=

    Text and patterns: For effective chain of thought, it takes two to tango , author=. arXiv preprint arXiv:2209.07686 , year=

    Pith/arXiv arXiv

  53. [53]

    arXiv preprint arXiv:2501.12948 , year=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    Pith/arXiv arXiv

  54. [54]

    arXiv preprint arXiv:2211.14275 , year=

    Solving math word problems with process-and outcome-based feedback , author=. arXiv preprint arXiv:2211.14275 , year=

    Pith/arXiv arXiv

  55. [55]

    arXiv preprint arXiv:2212.08073 , year=

    Constitutional ai: Harmlessness from ai feedback , author=. arXiv preprint arXiv:2212.08073 , year=

    Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2502.05167 , year=

    NoLiMa: Long-Context Evaluation Beyond Literal Matching , author=. arXiv preprint arXiv:2502.05167 , year=

    Pith/arXiv arXiv

  57. [57]

    5: Scaling reinforcement learning with llms , author=

    Kimi k1. 5: Scaling reinforcement learning with llms , author=. arXiv preprint arXiv:2501.12599 , year=

    Pith/arXiv arXiv

  58. [58]

    arXiv preprint arXiv:2305.20050 , year=

    Let's verify step by step , author=. arXiv preprint arXiv:2305.20050 , year=

    Pith/arXiv arXiv

  59. [59]

    arXiv preprint arXiv:2406.06592 , year=

    Improve Mathematical Reasoning in Language Models by Automated Process Supervision , author=. arXiv preprint arXiv:2406.06592 , year=

    Pith/arXiv arXiv

  60. [60]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  61. [61]

    Haosheng Zou, Xiaowei Lv, Shousheng Jia and Xiangzheng Zhang , title =

  62. [62]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. ICLR , volume=

  63. [63]

    arXiv preprint arXiv:2410.12832 , year=

    Generative reward models , author=. arXiv preprint arXiv:2410.12832 , year=

    Pith/arXiv arXiv

  64. [64]

    arXiv preprint arXiv:2408.15240 , year=

    Generative verifiers: Reward modeling as next-token prediction , author=. arXiv preprint arXiv:2408.15240 , year=

    Pith/arXiv arXiv

  65. [65]

    arXiv preprint arXiv:2504.02495 , year=

    Inference-time scaling for generalist reward modeling , author=. arXiv preprint arXiv:2504.02495 , year=

    arXiv

  66. [66]

    arXiv preprint arXiv:2112.09332 , year=

    Webgpt: Browser-assisted question-answering with human feedback , author=. arXiv preprint arXiv:2112.09332 , year=

    Pith/arXiv arXiv

  67. [67]

    Advances in Neural Information Processing Systems , volume=

    Webshop: Towards scalable real-world web interaction with grounded language agents , author=. Advances in Neural Information Processing Systems , volume=

  68. [68]

    Advances in Neural Information Processing Systems , volume=

    Mind2web: Towards a generalist agent for the web , author=. Advances in Neural Information Processing Systems , volume=

  69. [69]

    arXiv preprint arXiv:2307.13854 , year=

    Webarena: A realistic web environment for building autonomous agents , author=. arXiv preprint arXiv:2307.13854 , year=

    Pith/arXiv arXiv

  70. [70]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Visualwebarena: Evaluating multimodal agents on realistic visual web tasks , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  71. [71]

    arXiv preprint arXiv:2504.12516 , year=

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

    Pith/arXiv arXiv

  72. [72]

    arXiv preprint arXiv:2508.05748 , year=

    Webwatcher: Breaking new frontier of vision-language deep research agent , author=. arXiv preprint arXiv:2508.05748 , year=

    Pith/arXiv arXiv

  73. [73]

    arXiv preprint arXiv:2602.02185 , year=

    Vision-deepresearch benchmark: Rethinking visual and textual search for multimodal large language models , author=. arXiv preprint arXiv:2602.02185 , year=

    arXiv

  74. [74]

    arXiv preprint arXiv:2210.03629 , year=

    React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year=

    Pith/arXiv arXiv

  75. [75]

    arXiv preprint arXiv:2602.02276 , year=

    Kimi K2.5: Visual Agentic Intelligence , author=. arXiv preprint arXiv:2602.02276 , year=

    Pith/arXiv arXiv

  76. [76]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  77. [77]

    arXiv preprint arXiv:2402.03300 , year=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  78. [78]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-vl technical report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  79. [79]

    Qwen3.5: Accelerating Productivity with Native Multimodal Agents , url =

    Qwen Team , month =. Qwen3.5: Accelerating Productivity with Native Multimodal Agents , url =

  80. [80]

    arXiv preprint arXiv:2512.02395 , year=

    Skywork-R1V4: Toward Agentic Multimodal Intelligence through Interleaved Thinking with Images and DeepResearch , author=. arXiv preprint arXiv:2512.02395 , year=

    arXiv

Showing first 80 references.