Pith. sign in

REVIEW 2 major objections 5 minor 45 references

Active visual agents fail catastrophically when given fewer interaction turns than they prefer; AdaTurn conditions the agent on the allowed turn budget and trains the final forced answer, raising four-turn accuracy on VisualProbe-Medium fro

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

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

T0 review · deepseek-v4-flash

2026-08-02 01:46 UTC pith:U2DRK63Q

load-bearing objection The core idea is right — active visual agents should know their turn budget — but the paper claims budget-conditioning without showing Tmax in any prompt, and that gap undercuts the central causal claim. the 2 major comments →

arxiv 2607.14547 v1 pith:U2DRK63Q submitted 2026-07-16 cs.CV

AdaTurn: Budget-Aware Test-Time Scaling for Active Visual Perception Agents

Susan Liang , Chao Huang , Filippos Bellos , Jing Bi , Jason J Corso , Chenliang Xu This is my paper
classification cs.CV
keywords active visual perceptiontest-time scalingbudget conditioningcatastrophic truncationforced-answer RLmulti-turn visual agentsrollout schedulingvisual grounding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper identifies catastrophic truncation as the central failure of deployment-constrained visual agents: when the rollout budget is smaller than the trajectory the agent implicitly expects, the interaction is cut off before any valid answer. AdaTurn makes the allowed number of turns an explicit part of the task, conditioning the policy on Tmax, and introduces Forced-Answer DAPO (FA-DAPO), which converts an over-budget tool request into a trainable final-answer step. The paper reports that this yields large low-budget gains—VisualProbe-Medium rises from 36.7% to 47.6% at four turns—while preserving high-budget scaling and transferring to other backbones. A reader should care because deployment budgets are rarely fixed, and current agents are trained for a single hidden rollout length.

Core claim

On its own terms, AdaTurn claims that a multi-turn visual agent can be calibrated to its deployment budget. The central discovery is that treating the budget boundary as a learnable decision—by rolling back an over-budget tool request and training the resulting forced answer with the normal accuracy reward—turns catastrophic truncation from a rollout-engine artifact into a supervised policy behavior. The paper shows consistent low-budget accuracy gains across VisualProbe, V* Bench, HR-Bench, and MME-RealWorld, no loss at 32 turns, and successful transfer to four- and eight-billion-parameter backbones.

What carries the argument

Forced-Answer DAPO (FA-DAPO): a reinforcement-learning procedure built on decoupled-clip DAPO that, when the policy would emit a tool call at the last permitted turn, rolls back to the last executable prefix, appends a budget-exhausted control message, and scores the resulting forced answer with the standard accuracy reward. This converts the over-budget event from a masked or penalized failure into a trainable final-decision step. It is complemented by the budget-conditioned agent loop—the state includes (Q, I, Tmax)—and by dynamic budget sampling with a load-balanced rollout scheduler that keeps worker utilization stable.

Load-bearing premise

The policy actually receives Tmax as a visible prompt input: the paper's formulation includes it in the state, but the released prompt templates show no budget line, so the only concrete budget signal is the post-hoc tool-rejection message; if Tmax never reaches the model's context, the budget-conditioning claim is unsupported and the gains must be credited to boundary training alone.

What would settle it

Run the published AdaTurn prompts on identical (image, question) pairs with Tmax=4 and Tmax=8 and compare the first tool-call distribution. If the prompts contain no budget phrase and the distributions are indistinguishable, the model cannot be budget-conditioned during search; alternatively, strip any budget mention from the system prompt and check whether the four-turn accuracy gain survives. The paper's Appendix F prompts appear to lack the Tmax line, so this is checkable directly.

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

If this is right

  • At four turns, AdaTurn lifts VisualProbe-Hard from 26.5% to 39.0%, VisualProbe-Medium from 36.7% to 47.6%, and MME-RealWorld from 50.1% to 64.0% relative to Mini-o3.
  • At thirty-two turns, AdaTurn matches or exceeds the same baseline on four of the five main benchmarks, showing the low-budget gains do not trade away long-budget scaling.
  • The recipe transfers to Qwen3-VL-4B and 8B backbones, producing the same kind of low-budget improvement.
  • The load-balanced scheduler reduces rollout time by 1.34x, making dynamic-budget training practical.

Where Pith is reading between the lines

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

  • If the budget-conditioning mechanism is real, similar conditioning should transfer to token-level thinking budgets or step budgets in text-only reasoners, not just visual crop actions.
  • The ablation (all AdaTurn variants beat Mini-o3 at four turns) suggests boundary training carries much of the gain; the paper does not isolate how much Tmax being visible in the prompt adds, since the published prompt templates contain no budget line.
  • A direct test: compare first-action distributions under Tmax=4 and Tmax=8 on identical inputs; identical distributions would mean the model is not compressing its search in advance.
  • The load-balanced scheduler is a general systems contribution for any dynamic-budget reinforcement-learning pipeline, independent of the visual-domain detail.

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

2 major / 5 minor

Summary. The paper proposes AdaTurn, a framework for training active visual perception agents that operate under variable rollout turn budgets. The key ideas are (i) conditioning the policy on the maximum allowed turns Tmax, (ii) a Forced-Answer DAPO (FA-DAPO) variant that converts over-budget tool requests into a trainable forced final-answer step, and (iii) dynamic-budget training with a load-balanced rollout scheduler. The experiments on VisualProbe, V* Bench, HR-Bench, and MME-RealWorld show substantial improvements in the low-budget regime (e.g., VisualProbe-Medium at 4 turns: 36.7% to 47.6%) while maintaining performance at 32 turns, and the approach transfers to Qwen3-VL-4B/8B backbones.

Significance. If the central claims hold, AdaTurn addresses a practically important deployment issue—catastrophic truncation when the available rollout budget is smaller than the policy's preferred trajectory. The reported gains are large and consistent across multiple benchmarks and backbone scales, and the paper includes a useful systems contribution (load-balanced rollout scheduler). The release of exact prompt templates and detailed ablation of the RL design are strengths. However, two issues currently undermine the attribution of the improvements to the claimed mechanisms: (1) the released prompts do not appear to contain Tmax despite the formalization claiming budget conditioning, and (2) the ablation does not include a test-time-only forced-answer baseline, so the contribution of FA-DAPO training vs. the inference-time forced-answer mechanism is not isolated.

major comments (2)
  1. [Section 3.1, Eq. (2) and Section 3.3 vs. Appendix F] The paper defines the policy input as x=(Q, I, Tmax) and states in Section 3.3 that Tmax is 'inject[ed] into the prompt as part of the agent state.' However, Appendix F, which explicitly says it provides 'the exact prompts used,' contains no turn-budget field in the system, tool-observation, tool-error, or force-to-answer templates. The only budget-related text is the rejection message that appears after the budget is exhausted. If Tmax is not visible during the rollout, the policy cannot compress its search in advance, and the scaling curves in Figure 6 would be explained entirely by the forced-answer boundary step. This is an internal inconsistency between the formalism and the released prompts. The authors should either show the prompt that actually contains Tmax or revise the formulation and the budget-conditioning claims.
  2. [Tables 1 and Figure 7] The central claim is that FA-DAPO—explicit training of the forced final answer—produces the low-budget improvement, not merely the forced-answer mechanism. The ablation in Figure 7 compares variants of the training objective (masking, final-turn-only, format reward), but it does not include the critical baseline in which the same test-time forced-answer prompt is applied to a policy without FA-DAPO training (e.g., Mini-o3 or a DAPO baseline). Without such a baseline, the gains could be largely due to the inference-time rollback-and-append procedure rather than to the training signal. Please add a 'Mini-o3 + forced-answer at test' or 'DAPO + forced-answer at test' condition to isolate the contribution of the FA-DAPO objective.
minor comments (5)
  1. [Figures 6 and 7] The x-axis is labeled '# Frames' but the paper is about turn budgets, not video frames. Rename to '# Turns' or 'Rollout Budget' for clarity.
  2. [Author affiliation] Typo in affiliation: 'Univeristy of Michigan' should be 'University of Michigan'.
  3. [Table 1] The DeepEyes baseline is reported at 6 turns. Please clarify whether 6 is the maximum budget used for that model and whether this is directly comparable to the 4-turn and 32-turn evaluations of other methods.
  4. [Appendix C] The phrase 'Masking prefix turns' is ambiguous. Specify exactly which tokens are masked and how this variant differs from the 'Final Turn RL' variant.
  5. [Appendix E] The Qwen3-32B judge is used both for RL rewards and for final evaluation. Since all baselines are evaluated with the same judge, relative comparisons are fair, but the absolute numbers may be sensitive to judge idiosyncrasies. Please state this limitation explicitly.

Circularity Check

0 steps flagged

No significant circularity; the reported gains are benchmark-measured against external baselines, with no equation or fitted parameter reducing to its inputs.

full rationale

AdaTurn is an empirical systems paper. Its central claims—low-budget accuracy gains, FA-DAPO as the main contributor, and scheduler speedup—are supported by measurements on external benchmarks and compared against externally reported Mini-o3/DeepEyes numbers. None of these results is derived from an equation that already contains the conclusion, and no fitted parameter is renamed as a prediction. The closest tension is the Tmax/prompt mismatch: Sec. 3.1 defines x=(Q,I,Tmax) and Sec. 3.3 says Tmax is 'inject[ed] into the prompt,' while Appendix F's exact prompts contain no explicit budget line. That is a real implementation/consistency gap that weakens the attribution of gains to budget conditioning, but it is not circular—the low-budget numbers are not recovered from the definition of Tmax. Similarly, using the same judge for reward and evaluation would be a measurement-protocol concern, not a definitional reduction. No load-bearing claim rests on a self-citation, and the paper makes no parameter-free derivation that could be equivalent to its inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

No invented entities. The free parameters are RL hyperparameters plus the training budget range; the key hidden assumptions are the judge's correctness and the visibility of Tmax. The paper contributes a training procedure, not a new physical or mathematical entity.

free parameters (4)
  • Training budget range [4,12] turns = 4 to 12
    Hand-picked; the model never trains beyond 12 turns yet is evaluated at 32, making high-budget results an extrapolation.
  • Reward weights lambda_acc, lambda_fmt = 1.0, 0.0
    Chosen in Table 3; ablation shows format reward hurts but no sensitivity analysis of the accuracy weight is reported.
  • DAPO clip thresholds epsilon_low, epsilon_high = not reported
    Required by Eq. (7); their values are not given, so the boundary RL objective is under-specified.
  • Group size G = 16
    RL group size from Table 3; affects advantage normalization but is not central to the contribution.
axioms (5)
  • standard math DAPO/GRPO clipped policy-gradient objective optimizes expected reward
    Eqs. (6)-(8) rely on standard policy-gradient theory; no convergence guarantee is given, but this is standard practice.
  • domain assumption Qwen3-32B judge provides correct binary labels
    Eq. (10) and all reported metrics use the judge in Appendix E; no human agreement or error analysis is reported.
  • ad hoc to paper Rolling back to the last prefix and appending the forced-answer prompt yields a valid learning signal
    Eq. (9) is the defining FA-DAPO transform; its effectiveness is assumed and tested only indirectly through ablations.
  • ad hoc to paper Tmax is available to the policy in its context
    Eqs. (1)-(2) assume budget conditioning, but Appendix F prompt templates do not contain Tmax; without this premise the central mechanism fails.
  • domain assumption Tool executor returns faithful zoom observations
    The agent's evidence is whatever E(I, y_t) returns; tool failures or low-quality crops would corrupt the forced answers.

pith-pipeline@v1.3.0-alltime-deepseek · 17772 in / 14156 out tokens · 143941 ms · 2026-08-02T01:46:07.118865+00:00 · methodology

0 comments
read the original abstract

Active visual agents solve fine-grained image tasks by interleaving reasoning with image-grounding actions across multiple turns. However, deployment-time rollout budgets are rarely fixed: some requests permit long rollouts, while others require the agent to act under a tight turn limit. Existing methods train the policy as if the rollout budget were hidden, so when the available budget is smaller than the trajectory the agent prefers, the interaction is often truncated before any valid answer is produced; we term this failure \emph{catastrophic truncation}. To overcome this challenge, we present AdaTurn, a budget-aware framework that conditions the agent on the allowed number of turns and explicitly trains the boundary behavior induced by the budget. Our key component, Forced-Answer DAPO (FA-DAPO), converts the over-budget event from a masked or penalized failure into a trainable final-decision step, teaching the model to synthesize partial evidence when further tool use is no longer possible. We further randomize rollout budgets during both training and inference and introduce a load-balanced scheduler that makes such operations practical. AdaTurn substantially improves low-budget accuracy, for example raising VisualProbe-Medium from 36.7% to 47.6% at four turns, while preserving strong scaling at larger budgets and transferring effectively to multiple backbones and general multimodal benchmarks.

Figures

Figures reproduced from arXiv: 2607.14547 by Chao Huang, Chenliang Xu, Filippos Bellos, Jason J Corso, Jing Bi, Susan Liang.

Figure 1
Figure 1. Figure 1: Turn-aware reasoning under limited rollout budgets. Existing visual agents can stop too [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of AdaTurn. The left panel shows the turn-aware agent loop: the model receives [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training behavior at the budget boundary. Prior methods either treat over-budget rollouts [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Load-balanced rollout assignment. Dynamic rollout budgets create large differences in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of load-balanced rollout scheduling. The left plot reports the minimum, maximum, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance vs. rollout budget. AdaTurn provides a more favorable trade-off than Mini [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: isolates the contribution of the reinforcement learning design. The main takeaway is that forced-answer training is the primary contributor to the improvement: every AdaTurn variant outperforms Mini-o3 at four turns, indicating that explicit supervision of the budget boundary is important for mitigating catastrophic truncation. Among these variants, the default formulation performs best overall, especially… view at source ↗
Figure 8
Figure 8. Figure 8: Successful rollout under a four-turn budget. The agent localizes the pedestrian in the yellow [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Successful rollout under a four-turn budget. The agent initially searches the fireplace region, [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Successful rollout under an eight-turn budget. The agent performs a longer sequence of [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Successful rollout under an eight-turn budget. The agent progressively zooms into a small [PITH_FULL_IMAGE:figures/full_fig_p021_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Failure case under a four-turn budget. The agent searches a crowded park scene for the [PITH_FULL_IMAGE:figures/full_fig_p022_12.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

45 extracted references · 14 linked inside Pith

  1. [1]

    Attention is All you Need , url =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, ukasz and Polosukhin, Illia , booktitle =. Attention is All you Need , url =

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Visual instruction tuning , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    Xu, Ruyi and Yao, Yuan and Guo, Zonghao and Cui, Junbo and Ni, Zanlin and Ge, Chunjiang and Chua, Tat-Seng and Liu, Zhiyuan and Sun, Maosong and Huang, Gao , journal=

  4. [4]

    Qwen2.5-

    Bai, Shuai and Chen, Keqin and Liu, Xuejing and Wang, Jialin and Ge, Wenbin and Song, Sibo and Dang, Kai and Wang, Peng and Wang, Shijie and Tang, Jun and others , journal=. Qwen2.5-

  5. [5]

    arXiv preprint arXiv:2412.05271 , year=

    Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling , author=. arXiv preprint arXiv:2412.05271 , year=

    Pith/arXiv arXiv

  6. [6]

    V*: Guided visual search as a core mechanism in multimodal

    Wu, Penghao and Xie, Saining , booktitle=. V*: Guided visual search as a core mechanism in multimodal

  7. [7]

    arXiv preprint arXiv:2408.15556 , year=

    Divide, conquer and combine: A training-free framework for high-resolution image perception in multimodal large language models , author=. arXiv preprint arXiv:2408.15556 , year=

    Pith/arXiv arXiv

  8. [8]

    Adaptive Chain-of-Focus reasoning via dynamic visual search and zooming for efficient

    Zhang, Xintong and Gao, Zhi and Zhang, Bofei and Li, Pengxiang and Zhang, Xiaowen and Liu, Yang and Yuan, Tao and Wu, Yuwei and Jia, Yunde and Zhu, Song-Chun and Li, Qing , journal=. Adaptive Chain-of-Focus reasoning via dynamic visual search and zooming for efficient

  9. [9]

    Zhang, Yi-Fan and Zhang, Huanyu and Tian, Haochen and Fu, Chaoyou and Zhang, Shuangqing and Wu, Junfei and Li, Feng and Wang, Kun and Wen, Qingsong and Zhang, Zhang and others , journal=

  10. [10]

    arXiv preprint arXiv:2509.07969 , year=

    Mini-o3: Scaling up reasoning patterns and interaction turns for visual search , author=. arXiv preprint arXiv:2509.07969 , year=

    Pith/arXiv arXiv

  11. [11]

    Zheng, Ziwei and Yang, Michael and Hong, Jack and Zhao, Chenxiao and Xu, Guohai and Yang, Le and Shen, Chao and Yu, Xing , journal=

  12. [12]

    Hong, Jack and Zhao, Chenxiao and Zhu, ChengLin and Lu, Weiheng and Xu, Guohai and Yu, Xing , journal=

  13. [13]

    Scaling agentic reinforcement learning for tool-integrated reasoning in

    Lu, Meng and Xu, Ran and Fang, Yi and Zhang, Wenxuan and Yu, Yue and Srivastava, Gaurav and Zhuang, Yuchen and Elhoseiny, Mohamed and Fleming, Charles and Yang, Carl and others , journal=. Scaling agentic reinforcement learning for tool-integrated reasoning in

  14. [14]

    Liu, Jiaqi and Xiong, Kaiwen and Xia, Peng and Zhou, Yiyang and Ji, Haonian and Feng, Lu and Han, Siwei and Ding, Mingyu and Yao, Huaxiu , journal=. Agent0-

  15. [15]

    Guo, Daya and Yang, Dejian and Zhang, He and Song, Junxiao and Zhang, Runxin and Xu, Runqi and Zhu, Qihao and Ma, Shirong and Wang, Peiyi and Bi, Xiao and others , journal=

  16. [16]

    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

  17. [17]

    arXiv preprint arXiv:2503.14476 , year=

    Dapo: An open-source llm reinforcement learning system at scale , author=. arXiv preprint arXiv:2503.14476 , year=

    Pith/arXiv arXiv

  18. [18]

    Shen, Haozhan and Liu, Peng and Li, Jingcheng and Fang, Chunxin and Ma, Yibo and Liao, Jiajia and Shen, Qiaoli and Zhang, Zilun and Zhao, Kangjia and Zhang, Qianqian and Xu, Ruochen and Zhao, Tiancheng , journal=

  19. [19]

    Zhang, Jingyi and Gao, Jiaxing and Pang, Yankai and Zhao, Ruyi and Wang, Xi and Zhang, Jingqun and Hou, Peng and Luo, Rong and Liu, Bin and Huang, Haibin , journal=

  20. [20]

    Visionary-

    Zhang, Jiaer and Li, Bingqi and Zhou, Wei and Zhao, Rui , journal=. Visionary-

  21. [21]

    Snell, Charlie and Lee, Jaehoon and Xu, Kelvin and Kumar, Aviral , journal=. Scaling

  22. [22]

    OpenAI Blog , year=

    Learning to reason with. OpenAI Blog , year=

  23. [23]

    arXiv preprint arXiv:2512.02008 , year=

    The art of scaling test-time compute for large language models , author=. arXiv preprint arXiv:2512.02008 , year=

    arXiv

  24. [24]

    arXiv preprint arXiv:2505.20522 , year=

    Scaling over scaling: Exploring test-time scaling plateau in large reasoning models , author=. arXiv preprint arXiv:2505.20522 , year=

    Pith/arXiv arXiv

  25. [25]

    Wen, Hao and Wu, Xinrui and Sun, Yi and Zhang, Feifei and Chen, Liye and Wang, Jie and Liu, Yunxin and Liu, Yunhao and Zhang, Ya-Qin and Li, Yuanchun , journal=

  26. [26]

    arXiv preprint arXiv:2509.07820 , year=

    Certainty-guided reasoning in large language models: A dynamic thinking budget approach , author=. arXiv preprint arXiv:2509.07820 , year=

    arXiv

  27. [27]

    arXiv preprint arXiv:2602.02276 , year=

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

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2408.03326 , year=

    Llava-onevision: Easy visual task transfer , author=. arXiv preprint arXiv:2408.03326 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2511.21631 , year=

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

    Pith/arXiv arXiv

  31. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    V?: Guided visual search as a core mechanism in multimodal llms , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  32. [32]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Dyfo: A training-free dynamic focus visual search for enhancing lmms in fine-grained visual understanding , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  33. [33]

    arXiv preprint arXiv:2505.15436 , year=

    Adaptive Chain-of-Focus Reasoning via Dynamic Visual Search and Zooming for Efficient VLMs , author=. arXiv preprint arXiv:2505.15436 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2505.15966 , year=

    Pixel reasoner: Incentivizing pixel-space reasoning with curiosity-driven reinforcement learning , author=. arXiv preprint arXiv:2505.15966 , year=

    Pith/arXiv arXiv

  35. [35]

    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

  36. [36]

    2024 , journal =

    HybridFlow: A Flexible and Efficient RLHF Framework , author =. 2024 , journal =

    2024

  37. [37]

    arXiv preprint arXiv:2411.16044 , year=

    ZoomEye: Enhancing Multimodal LLMs with Human-Like Zooming Capabilities through Tree-Based Image Exploration , author=. arXiv preprint arXiv:2411.16044 , year=

    Pith/arXiv arXiv

  38. [38]

    The 36th Conference on Neural Information Processing Systems , year=

    Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering , author=. The 36th Conference on Neural Information Processing Systems , year=

  39. [39]

    arXiv preprint arXiv:2310.02255 , year=

    MathVista: Evaluating Math Reasoning in Visual Contexts with GPT-4V, Bard, and Other Large Multimodal Models , author=. arXiv preprint arXiv:2310.02255 , year=

    Pith/arXiv arXiv

  40. [40]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , year =

    Masry, Ahmed and Long, Do Xuan and Tan, Jia Qing and Joty, Shafiq and Hoque, Enamul , title =. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics , year =

  41. [41]

    Mathew, Minesh and Karatzas, Dimosthenis and Jawahar, C. V. , title =. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , year =

  42. [42]

    2025 , eprint=

    SimpleTIR: End-to-End Reinforcement Learning for Multi-Turn Tool-Integrated Reasoning , author=. 2025 , eprint=

    2025

  43. [43]

    2025 , eprint=

    Group Sequence Policy Optimization , author=. 2025 , eprint=

    2025

  44. [44]

    Science China Information Sciences , volume=

    Ocrbench: on the hidden mystery of ocr in large multimodal models , author=. Science China Information Sciences , volume=. 2024 , publisher=

    2024

  45. [45]

    Advances in Neural Information Processing Systems , volume=

    Cambrian-1: A fully open, vision-centric exploration of multimodal llms , author=. Advances in Neural Information Processing Systems , volume=