pith. sign in

arxiv: 2606.31504 · v1 · pith:JRAGTKTDnew · submitted 2026-06-30 · 💻 cs.CV

SimpleSearch-VL: A Simple Recipe for Multimodal Agentic Deep Search

Ming Dai , Zhihong Lu , Jinjie Gu , Jiedong Zhuang , Yefeng Liu , Wankou Yang , Jian Wang , Chunhua Shen This is my paper

Pith reviewed 2026-07-01 05:59 UTC · model grok-4.3

classification 💻 cs.CV
keywords multimodal agentsagentic searchvisual language modelstool usereinforcement learningchain of thought verificationrollout efficiency
0
0 comments X

The pith

A lightweight training recipe using 7K trajectories lets open multimodal agents match Gemini-3-Pro on search tasks.

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

The paper introduces SimpleSearch-VL to strengthen an agent's own search-and-verification loop instead of scaling data, tools, or helper models. It applies Factorized Adaptive Rollout to create more useful training groups and reuse samples against long-tail delays, adds explicit chain-of-thought checks that verify whether retrieved visual and text cues actually address the query, and keeps all processing inside the agent with a simple self-summary step. These three changes, trained on only 5K supervised tool-interleaved examples plus 2K RL steps, raise Qwen3-VL agent baselines by roughly 16 points on average, bringing the 30B variant to parity with agentic Gemini-3-Pro.

Core claim

The central claim is that Factorized Adaptive Rollout, evidence-verified reasoning, and in-agent webpage self-summary together allow an agent trained on 5K supervised tool-interleaved trajectories and 2K RL data to improve Qwen3-VL agentic baselines by 15.8 and 16.0 average points for the 8B and 30B-A3B variants, with the larger model reaching performance competitive with agentic Gemini-3-Pro.

What carries the argument

Factorized Adaptive Rollout (FAR) that forms informative training groups while recycling redundant samples to reduce latency and surface hard cases, paired with evidence-verified chain-of-thought reasoning that scores the relevance of each retrieved visual and textual cue to the original query.

If this is right

  • Multimodal agents can reach competitive search performance without external auxiliary models or large additional datasets.
  • Verification steps inserted into the reasoning chain measurably reduce errors when agents must judge retrieved images and text.
  • Keeping summary generation inside the agent removes the need for separate summarization services.
  • Adaptive rollout strategies improve sample efficiency even when total training data stays small.
  • The same three components can be applied to other base vision-language models without architectural changes.

Where Pith is reading between the lines

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

  • If the gains transfer to other open VLMs, the recipe offers a low-cost path to stronger agentic search across model families.
  • The emphasis on internal verification may extend to non-search agent tasks where agents must judge the utility of retrieved evidence.
  • Lightweight self-contained agents could lower deployment costs for applications that currently rely on closed-model APIs.
  • Testing FAR on longer-horizon tasks would reveal whether the latency-mitigation benefit scales beyond the reported search benchmarks.

Load-bearing premise

The reported gains come from the proposed rollout, verification, and self-summary steps rather than from hidden differences in evaluation setup, data selection, or how the baselines were run.

What would settle it

Reproduce the exact Qwen3-VL agent baselines and SimpleSearch-VL models on the same test sets with identical tool interfaces and prompt templates to check whether the 15-plus-point average gains remain.

Figures

Figures reproduced from arXiv: 2606.31504 by Chunhua Shen, Jian Wang, Jiedong Zhuang, Jinjie Gu, Ming Dai, Wankou Yang, Yefeng Liu, Zhihong Lu.

Figure 1
Figure 1. Figure 1: Performance comparison with representative multimodal deep search agents. SimpleSearch￾VL-8B outperforms most open-source 30B-scale multimodal deep search agents, while SimpleSearch-VL￾30B-A3B achieves performance competitive with agentic Gemini-3-Pro. *Corresponding authors: Jian Wang (bobblair.wj@antgroup.com) and Wankou Yang (wkyang@seu.edu.cn). 1 arXiv:2606.31504v1 [cs.CV] 30 Jun 2026 [PITH_FULL_IMAGE… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Factorized Adaptive Rollout. (a)–(d) compare fixed sampling, prompt expansion, rollout allocation, and FAR, which keeps exploring hard groups while skipping redundant tail rollouts. (e)–(f) show rollout time and the accuracy–efficiency trade-off. 1 Introduction Agentic search extends large language models from passive response generation to active evidence acquisition. In the language-only sett… view at source ↗
Figure 3
Figure 3. Figure 3: Agentic search process of SimpleSearch-VL. The model alternates between reasoning, tool calls, evidence verification, and final answer generation. Here ai is a previous action and oi is the corresponding tool observation. Conditioning on Ht makes the state contain all information needed for the next decision. • Action at ∈ A: The policy first generates <thinking> and then emits exactly one action: at ∼ πθ … view at source ↗
Figure 4
Figure 4. Figure 4: FAR sampling details and filtering process. The numbers 1, 2, 3, and subsequent indices indicate the intuitive sampling order. FAR fills base rollouts first, then allocates expanded rollout slots horizontally across groups. Once an answer-correct rollout appears for a group, later Rollout Allocation slots for that group are filtered, avoiding redundant rollouts. Only groups satisfying Eq. 4 are retained fo… view at source ↗
Figure 5
Figure 5. Figure 5: a, Prompt Expansion improves performance by adding more effective samples within each training step, thereby accelerating convergence. Fig. 5b supports this trend: larger expansion ratios provide more signals in the early stage, but the gain becomes marginal later because most prompt groups become fully correct and no longer provide useful reward variation. Fig. 5c shows that rollout time increases with th… view at source ↗
Figure 6
Figure 6. Figure 6: Fixed rollout-budget trade-off. (a) Best evaluation scores for each rollout budget. (b) Signal rate during training. (c) Per-step rollout time. Method Recovered Groups Rate Rollout 4–6 117 11.7% Rollout 4–8 195 17.7% (a) Recovery metrics. 0 20 40 60 80 100 5 10 15 20 All-Wrong Groups Rollout 4--6 Rollout 4--8 (b) All-wrong groups. 0 20 40 60 80 100 0 3 6 Recovered Groups Rollout 4--6 Rollout 4--8 (c) Recov… view at source ↗
Figure 7
Figure 7. Figure 7: Hard-group recovery with additional rollouts. (a) Recovered occurrences and recovery rate among hard groups. (b) Initially all-wrong group occurrences per step. (c) Occurrences recovered by later rollouts. Method Step(s) AVG MMS BC-VL FVQA Standard 192 56.6 66.8 35.0 68.0 Rollout 4–6 118 56.8 65.4 37.0 68.0 Rollout 4–8 134 60.4 67.2 41.0 73.0 Prompt 2× 273 59.7 67.2 43.0 69.0 FAR (2×,4–6) 165 62.0 73.0 45.… view at source ↗
Figure 8
Figure 8. Figure 8: FAR accuracy–efficiency trade-off. (a) Rollout cost and evaluation scores. (b) Signal rate. (c) All-correct rate. Factorized Adaptive Rollout. Fig. 8a shows that expanding the rollout budget through Rollout Allocation brings clear performance gains. We do not emphasize the reward curve in this setting because the average reward becomes less comparable under adaptive sampling: Rollout Allocation continues s… view at source ↗
Figure 9
Figure 9. Figure 9: compares SFT+RL and Direct RL under the same FAR (2×, 4–6) setting. The SFT-initialized run keeps a longer and more stable search process, with average reasoning rounds changing only from 5.6 to 5.4 between steps 10 and 100. Direct RL starts from shorter trajectories and increases from 3.4 to 4.2 rounds, indicating that it is still learning when to search and verify evidence. The sample-level image￾search … view at source ↗
Figure 10
Figure 10. Figure 10: Inference rounds and Pass@1. Bars show the average number of reasoning rounds on each benchmark, while markers and dashed segments report the corresponding Pass@1 scores. 4.3.6 Training Cost Comparison [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Raw source composition of SFT and RL training data. (a) SFT trajectory mixture. (b) RL data mixture. 0 20 40 60 80 100 Share of training samples (%) SFT RL 4,827 (93.0%) 366 (7.0%) 1,592 (79.8%) 403 (20.2%) (a) Multi-image Ratio Single-image Multi-image 0 1 2 3 4 5 6 7 8 9 10 Tool-use rounds before final answer 0 10 20 30 40 50 SFT samples (%) 37 46.8% 14.7% 21.8% 4.9% 8.4% 57 34 Avg. = 3.20 (b) SFT Round… view at source ↗
Figure 12
Figure 12. Figure 12: Input-image type and SFT round distribution. (a) Single-image and multi-image shares in SFT and RL data. (b) SFT tool-use round distribution, where a round denotes one tool call before the final answer. B Agent Workflow and Tool Interface The agent operates in a tool-interleaved loop. Given a question and one or more input images, the system assigns explicit image identifiers, e.g., img_idx=0 and img_idx=… view at source ↗
Figure 13
Figure 13. Figure 13: Standard single-image case. The agent searches the left image region, matches retrieved thumbnails and titles to Don Lemon, and verifies the identity with visited webpages before answering. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Standard single-image case with visual-to-text verification. The reverse-image-search result grounds the scene as the Evil Queen from Snow White; webpage and text evidence then confirm Gal Gadot as the actress in the new film. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Single-image case without image search. For this straightforward logo-based question, the model does not call image search; it uses the visible Fine Art America text as an anchor and verifies the answer through text search and webpage visiting. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Multi-frame case with frame selection. The first cockpit frame gives generic simulator evidence, while the second frame retrieves a United Airlines training video; the agent then verifies that the facility is the United Airlines Flight Training Center in Denver. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Multi-frame case with cross-image comparison. The agent links the first frame to the 1932 Ford Hunger March and the second frame to the 1941 Ford Strikers Riot, verifies both dates through webpage evidence, and computes the correct nine-year interval. 33 [PITH_FULL_IMAGE:figures/full_fig_p033_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Multi-region case. After a broad whole-image reverse search, the agent searches localized player regions, uses the goalkeeper match to anchor the England-squad context, and verifies from webpages that Thomas Tuchel recalled Jordan Henderson. 34 [PITH_FULL_IMAGE:figures/full_fig_p034_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Failure case with wrong entity selection. A generic songwriting image leads the agent to search for a plausible Songwriters Hall of Fame induction and answer Taylor Swift, although the benchmark target is Ben Peters. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Failure case after correct visual grounding. Reverse image search identifies Dan Reeder, but the later track-list comparison selects “Gin Tonic” instead of the target shared song, “Fun Campfire Song”. 37 [PITH_FULL_IMAGE:figures/full_fig_p037_20.png] view at source ↗
read the original abstract

We present SimpleSearch-VL, an efficient, reliable, and practical framework for multimodal agentic search. Its core idea is to improve the agent's own search-and-verification process rather than scaling data, tools, or auxiliary model components. For efficiency, Factorized Adaptive Rollout (FAR) improves sampling efficiency by forming more informative training groups while using redundant samples to mitigate long-tail latency and expose hard samples. For reliability, SimpleSearch-VL performs evidence-verified reasoning, explicitly using chain-of-thought verification to assess the relevance of retrieved visual and textual cues to the original context. For practicality, SimpleSearch-VL keeps a lightweight tool interface and performs webpage self-summary within the agent, requiring no additional external dependencies. With only 5K supervised tool-interleaved trajectories and 2K RL data, SimpleSearch-VL improves Qwen3-VL agentic baselines by 15.8 and 16.0 average points for the 8B and 30B-A3B variants, respectively. The SimpleSearch-VL-30B-A3B model further achieves performance competitive with agentic Gemini-3-Pro.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper presents SimpleSearch-VL, a multimodal agentic search framework whose core contributions are Factorized Adaptive Rollout (FAR) for sampling efficiency, evidence-verified chain-of-thought reasoning for reliability, and in-agent webpage self-summary for practicality. Using only 5K supervised tool-interleaved trajectories and 2K RL samples, the method is reported to improve Qwen3-VL agentic baselines by 15.8 and 16.0 average points on the 8B and 30B-A3B variants respectively, with the larger model reaching performance competitive with agentic Gemini-3-Pro.

Significance. If the reported gains are shown to be robustly attributable to the three proposed components under identical evaluation conditions, the work would offer a practical, low-data recipe for improving multimodal agents without auxiliary models or external dependencies. The emphasis on efficiency via FAR and internal self-summary is a clear strength for deployment-oriented research.

major comments (2)
  1. [Abstract] Abstract: The central claim attributes the 15.8/16.0-point lifts to FAR, evidence-verified reasoning, and lightweight self-summary, yet provides no information on whether the Qwen3-VL baselines were re-implemented with identical tool interfaces, prompt formats, retrieval corpora, and scoring rules. Without this, the attribution cannot be verified and the numerical improvements do not yet support the causal claim.
  2. [Experimental section] Experimental section (assumed §4 or equivalent): No details are supplied on statistical significance, variance across runs, or controls for post-hoc data selection. The soundness of the 15.8/16.0-point gains therefore cannot be assessed from the provided information.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address the concerns about baseline equivalence and experimental robustness below, and will revise the manuscript to improve clarity on these points.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim attributes the 15.8/16.0-point lifts to FAR, evidence-verified reasoning, and lightweight self-summary, yet provides no information on whether the Qwen3-VL baselines were re-implemented with identical tool interfaces, prompt formats, retrieval corpora, and scoring rules. Without this, the attribution cannot be verified and the numerical improvements do not yet support the causal claim.

    Authors: We re-implemented the Qwen3-VL baselines using identical tool interfaces, prompt formats, retrieval corpora, and scoring rules to ensure fair attribution of gains to the proposed components (FAR, evidence-verified reasoning, and self-summary). This equivalence is stated in the experimental setup. We will add an explicit clarifying sentence to the abstract and expand the relevant paragraph in the introduction. revision: yes

  2. Referee: [Experimental section] Experimental section (assumed §4 or equivalent): No details are supplied on statistical significance, variance across runs, or controls for post-hoc data selection. The soundness of the 15.8/16.0-point gains therefore cannot be assessed from the provided information.

    Authors: The training trajectories were fixed before evaluation with no post-hoc selection. Main results reflect averages over three independent runs with different seeds, yielding consistent gains. We will add a dedicated paragraph in the experimental section reporting the number of runs, observed variance where available, and confirmation of the fixed data protocol. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance claims rest on reported training data and baselines without reduction to fitted inputs or self-citations.

full rationale

The paper presents an empirical framework using supervised trajectories and RL data to train a multimodal agent, claiming gains from components like FAR and evidence-verified reasoning. No equations, fitted parameters renamed as predictions, or self-citation chains appear in the provided text. The central claims concern measured improvements over baselines rather than any derivation that reduces by construction to its inputs. This is the standard case of a self-contained empirical report whose validity depends on experimental controls, not definitional equivalence.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no identifiable free parameters, axioms, or invented entities; full methods would be required to populate the ledger.

pith-pipeline@v0.9.1-grok · 5753 in / 1290 out tokens · 29205 ms · 2026-07-01T05:59:57.437095+00:00 · methodology

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

112 extracted references · 64 canonical work pages · 28 internal anchors

  1. [1]

    arXiv preprint arXiv:2505.22019 , year=

    VRAG-RL: Empower Vision-Perception-Based RAG for Visually Rich Information Understanding via Iterative Reasoning with Reinforcement Learning , author=. arXiv preprint arXiv:2505.22019 , year=

    work page arXiv

  2. [2]

    Thinking with Images

    DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning , author=. 2026 , eprint=

    2026

  3. [3]

    Wikipedia , howpublished=

  4. [4]

    Team, Anthropic , title=

  5. [5]

    2025 , eprint=

    A Survey on LLM-as-a-Judge , author=. 2025 , eprint=

    2025

  6. [6]

    arXiv preprint arXiv:2504.07956 , year=

    Vcr-bench: A comprehensive evaluation framework for video chain-of-thought reasoning , author=. arXiv preprint arXiv:2504.07956 , year=

    work page arXiv

  7. [7]

    2024 , eprint=

    GPT-4o System Card , author=. 2024 , eprint=

    2024

  8. [8]

    arXiv preprint arXiv:2503.17736 , year=

    V2P-Bench: Evaluating Video-Language Understanding with Visual Prompts for Better Human-Model Interaction , author=. arXiv preprint arXiv:2503.17736 , year=

    work page arXiv

  9. [9]

    arXiv preprint arXiv:2510.01304 , year=

    Agentic Jigsaw Interaction Learning for Enhancing Visual Perception and Reasoning in Vision-Language Models , author=. arXiv preprint arXiv:2510.01304 , year=

    work page arXiv

  10. [10]

    Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models

    Vision-r1: Incentivizing reasoning capability in multimodal large language models , author=. arXiv preprint arXiv:2503.06749 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents

    Visrag: Vision-based retrieval-augmented generation on multi-modality documents , author=. arXiv preprint arXiv:2410.10594 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Open-domain visual entity recognition: Towards recognizing millions of wikipedia entities , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  13. [13]

    European Conference on Computer Vision , pages=

    Sharegpt4v: Improving large multi-modal models with better captions , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Sharegpt4video: Improving video understanding and generation with better captions , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Are we on the right way for evaluating large vision-language models? , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Mmsearch: Unveiling the potential of large models as multi-modal search engines , author=

  17. [17]

    arXiv preprint arXiv:2601.03193 , year=

    UniCorn: Towards Self-Improving Unified Multimodal Models through Self-Generated Supervision , author=. arXiv preprint arXiv:2601.03193 , year=

    work page arXiv

  18. [18]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Enhancing Large Vision-Language Models with Ultra-Detailed Image Caption Generation , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  19. [19]

    arXiv preprint arXiv:2511.22134 , year=

    DualVLA: Building a Generalizable Embodied Agent via Partial Decoupling of Reasoning and Action , author=. arXiv preprint arXiv:2511.22134 , year=

    work page arXiv

  20. [20]

    arXiv preprint arXiv:2506.13977 , year=

    CRITICTOOL: Evaluating Self-Critique Capabilities of Large Language Models in Tool-Calling Error Scenarios , author=. arXiv preprint arXiv:2506.13977 , year=

    work page arXiv

  21. [21]

    arXiv preprint arXiv:2509.06945 , year=

    Interleaving reasoning for better text-to-image generation , author=. arXiv preprint arXiv:2509.06945 , year=

    work page arXiv

  22. [22]

    arXiv preprint arXiv:2504.05288 , year=

    LiveVQA: Live Visual Knowledge Seeking , author=. arXiv preprint arXiv:2504.05288 , year=

    work page arXiv

  23. [23]

    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=

    work page arXiv

  24. [24]

    arXiv preprint arXiv:2504.07165 , year=

    Perception in reflection , author=. arXiv preprint arXiv:2504.07165 , year=

    work page arXiv

  25. [25]

    2025 , eprint=

    PALADIN: Self-Correcting Language Model Agents to Cure Tool-Failure Cases , author=. 2025 , eprint=

    2025

  26. [26]

    2025 , eprint=

    Where LLM Agents Fail and How They can Learn From Failures , author=. 2025 , eprint=

    2025

  27. [27]

    Neurocomputing , pages=

    RE-GRPO: Leveraging hard negative cases through large language model guided self training , author=. Neurocomputing , pages=. 2025 , publisher=

    2025

  28. [28]

    2026 , eprint=

    On Group Relative Policy Optimization Collapse in Agent Search: The Lazy Likelihood-Displacement , author=. 2026 , eprint=

    2026

  29. [29]

    2026 , eprint=

    When Right Meets Wrong: Bilateral Context Conditioning with Reward-Confidence Correction for GRPO , author=. 2026 , eprint=

    2026

  30. [30]

    2026 , eprint=

    CoRPO: Adding a Correctness Bias to GRPO Improves Generalization , author=. 2026 , eprint=

    2026

  31. [31]

    2026 , eprint=

    WS-GRPO: Weakly-Supervised Group-Relative Policy Optimization for Rollout-Efficient Reasoning , author=. 2026 , eprint=

    2026

  32. [32]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  33. [33]

    arXiv preprint arXiv:2602.13179 , year=

    Fix Before Search: Benchmarking Agentic Query Visual Pre-processing in Multimodal Retrieval-augmented Generation , author=. arXiv preprint arXiv:2602.13179 , year=

    work page arXiv

  34. [34]

    arXiv preprint arXiv:2603.01050 , year=

    Mm-deepresearch: A simple and effective multimodal agentic search baseline , author=. arXiv preprint arXiv:2603.01050 , year=

    work page arXiv

  35. [35]

    2025 , eprint=

    Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning , author=. 2025 , eprint=

    2025

  36. [36]

    MMSearch-R1: Incentivizing LMMs to Search

    MMSearch-R1: Incentivizing LMMs to Search , author=. arXiv preprint arXiv:2506.20670 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    2026 , eprint=

    OpenVLThinkerV2: A Generalist Multimodal Reasoning Model for Multi-domain Visual Tasks , author=. 2026 , eprint=

    2026

  38. [38]

    2026 , eprint=

    Kimi K2.5: Visual Agentic Intelligence , author=. 2026 , eprint=

    2026

  39. [39]

    arXiv preprint arXiv:2510.12801 , year=

    Deepmmsearch-r1: Empowering multimodal llms in multimodal web search , author=. arXiv preprint arXiv:2510.12801 , year=

    work page arXiv

  40. [40]

    WebWatcher: Breaking New Frontier of Vision-Language Deep Research Agent

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

    work page internal anchor Pith review Pith/arXiv arXiv

  41. [41]

    DeepEyesV2: Toward Agentic Multimodal Model

    DeepEyesV2: Toward Agentic Multimodal Model , author=. arXiv preprint arXiv:2511.05271 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  42. [42]

    Tongyi DeepResearch Technical Report

    Tongyi deepresearch technical report , author=. arXiv preprint arXiv:2510.24701 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  43. [43]

    arXiv preprint arXiv:2505.22648 , year=

    Webdancer: Towards autonomous information seeking agency , author=. arXiv preprint arXiv:2505.22648 , year=

    work page arXiv

  44. [44]

    WebSailor: Navigating Super-human Reasoning for Web Agent

    WebSailor: Navigating Super-human Reasoning for Web Agent , author=. arXiv preprint arXiv:2507.02592 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2507.15061 , year=

    Webshaper: Agentically data synthesizing via information-seeking formalization , author=. arXiv preprint arXiv:2507.15061 , year=

    work page arXiv

  46. [46]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Simplevqa: Multimodal factuality evaluation for multimodal large language models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  47. [47]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Fvqa: Fact-based visual question answering , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2017 , publisher=

    2017

  48. [48]

    arXiv preprint arXiv:2302.11713 , year=

    Can pre-trained vision and language models answer visual information-seeking questions? , author=. arXiv preprint arXiv:2302.11713 , year=

    work page arXiv

  49. [49]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

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

    work page internal anchor Pith review Pith/arXiv arXiv

  50. [50]

    Proximal Policy Optimization Algorithms

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  51. [51]

    Advances in neural information processing systems , volume=

    Visual instruction tuning , author=. Advances in neural information processing systems , volume=

  52. [52]

    The eleventh international conference on learning representations , year=

    React: Synergizing reasoning and acting in language models , author=. The eleventh international conference on learning representations , year=

  53. [53]

    arXiv preprint arXiv:2501.07572 , year=

    Webwalker: Benchmarking llms in web traversal , author=. arXiv preprint arXiv:2501.07572 , year=

    work page arXiv

  54. [54]

    2022 , eprint=

    WebQA: Multihop and Multimodal QA , author=. 2022 , eprint=

    2022

  55. [55]

    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

  56. [56]

    2025 , note=

    rLLM: A Framework for Post-Training Language Agents , author=. 2025 , note=

    2025

  57. [57]

    Nature , volume=

    DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning , author=. Nature , volume=. 2025 , publisher=

    2025

  58. [58]

    Deepswe: Training a fully open-sourced, state-of-the-art coding agent by scaling rl, Jul 2025 , author=

    2025

  59. [59]

    AdaTooler-V: Adaptive Tool-Use for Images and Videos

    AdaTooler-V: Adaptive Tool-Use for Images and Videos , author=. arXiv preprint arXiv:2512.16918 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  60. [60]

    arXiv preprint arXiv:2502.01600 , year=

    Reinforcement learning for long-horizon interactive llm agents , author=. arXiv preprint arXiv:2502.01600 , year=

    work page arXiv

  61. [61]

    Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs

    Back to basics: Revisiting reinforce style optimization for learning from human feedback in llms , author=. arXiv preprint arXiv:2402.14740 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  62. [62]

    arXiv preprint arXiv:2510.26788 , year=

    Defeating the training-inference mismatch via fp16 , author=. arXiv preprint arXiv:2510.26788 , year=

    work page arXiv

  63. [63]

    arXiv preprint arXiv:2508.21475 , year=

    MMSearch-Plus: Benchmarking Provenance-Aware Search for Multimodal Browsing Agents , author=. arXiv preprint arXiv:2508.21475 , year=

    work page arXiv

  64. [64]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  65. [65]

    Video-R1: Reinforcing Video Reasoning in MLLMs

    Video-r1: Reinforcing video reasoning in mllms , author=. arXiv preprint arXiv:2503.21776 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  66. [66]

    OpenAI GPT-5 System Card

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

    work page internal anchor Pith review Pith/arXiv arXiv

  67. [67]

    preprint , year=

    Vision-DeepResearch Benchmark: Rethinking Visual and Textual Search for Multimodal Large Language Models , author=. preprint , year=

  68. [68]

    Exploring Reasoning Reward Model for Agents

    Exploring Reasoning Reward Model for Agents , author=. arXiv preprint arXiv:2601.22154 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  69. [69]

    arXiv preprint arXiv:2506.04207 , year=

    Advancing Multimodal Reasoning: From Optimized Cold Start to Staged Reinforcement Learning , author=. arXiv preprint arXiv:2506.04207 , year=

    work page arXiv

  70. [70]

    arXiv preprint arXiv:2510.08457 , year=

    ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping , author=. arXiv preprint arXiv:2510.08457 , year=

    work page arXiv

  71. [71]

    arXiv preprint arXiv:2603.29620 , year=

    Unify-Agent: A Unified Multimodal Agent for World-Grounded Image Synthesis , author=. arXiv preprint arXiv:2603.29620 , year=

    work page arXiv

  72. [72]

    Gen-Searcher: Reinforcing Agentic Search for Image Generation

    Gen-Searcher: Reinforcing Agentic Search for Image Generation , author=. arXiv preprint arXiv:2603.28767 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  73. [73]

    arXiv preprint arXiv:2601.22060 , year=

    Vision-deepresearch: Incentivizing deepresearch capability in multimodal large language models , author=. arXiv preprint arXiv:2601.22060 , year=

    work page arXiv

  74. [74]

    OneThinker: All-in-one Reasoning Model for Image and Video

    Onethinker: All-in-one reasoning model for image and video , author=. arXiv preprint arXiv:2512.03043 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  75. [75]

    2026 , eprint=

    AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning , author=. 2026 , eprint=

    2026

  76. [76]

    Qwen3-VL Technical Report

    Qwen3-VL Technical Report , author=. arXiv preprint arXiv:2511.21631 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  77. [77]

    2025 , eprint=

    Agentic Reinforced Policy Optimization , author=. 2025 , eprint=

    2025

  78. [78]

    2025 , eprint=

    WebSailor: Navigating Super-human Reasoning for Web Agent , author=. 2025 , eprint=

    2025

  79. [79]

    arXiv preprint arXiv:2505.14246 , year=

    Visual Agentic Reinforcement Fine-Tuning , author=. arXiv preprint arXiv:2505.14246 , year=

    work page arXiv

  80. [80]

    arXiv preprint arXiv:2512.24330 , year=

    SenseNova-MARS: Empowering Multimodal Agentic Reasoning and Search via Reinforcement Learning , author=. arXiv preprint arXiv:2512.24330 , year=

    work page arXiv

Showing first 80 references.