Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

TACO: Think-Answer Consistency for Optimized Long-Chain Reasoning and Efficient Data Learning via Reinforcement Learning in LVLMs

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Tying RL rewards to reasoning lifts OOD vision-language scores by 20

desk verdict Credible RL training recipe for LVLMs with consistent gains, but the headline OOD numbers rest on an under-specified test-time ensemble that needs clarification before the numbers can be trusted. read the letter →

arxiv 2505.20777 v1 pith:Z6FP4LKY submitted 2025-05-27 cs.CV

Zhehan Kan , Yanlin Liu , Kun Yin , Xinghua Jiang , Xin Li , Haoyu Cao , Yinsong Liu , Deqiang Jiang
show 3 more authors
Xing Sun Qingmin Liao Wenming Yang
This is my paper
classification cs.CV
keywords reinforcementlearningvision-languagemodelsvisualreasoningreferringexpressioncomprehensionquestionansweringGRPOtest-timeresolutionscalingthink-answerconsistency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that vision-language models can be trained with reinforcement learning to reason first and answer second, provided the reward checks both steps against the ground truth. The proposed algorithm, TACO, couples the model's thinking output and its final box or text answer into a single consistency-aware reward, and adds three stabilizers: it rolls back samples whose policy drift is too large, re-weights sampling toward moderate-difficulty examples, and rescales test images across multiple resolutions. On referring-expression comprehension the method reports gains of about four points over the base model in-domain and roughly twenty points on an out-of-domain reasoning-grounding benchmark, with smaller but consistent gains on a suite of VQA tasks. The paper's point is that long-chain reasoning in multimodal models is learnable and stable if the reward architecture enforces a tight think-answer link.

What carries the argument

The load-bearing mechanism is the Think-Answer Consistency reward, a reward that measures agreement among the model's reasoning box, its final answer box, and the ground-truth box via three-way IoU, and, in VQA, via a supervisor-assessed semantic consistency score. Around this reward sit three stabilizers: Rollback Resample Strategy, which computes per-sample KL divergence between current and reference policies and masks gradients of samples above a threshold while down-weighting their sampling probability; Adaptive Difficulty Sampling, which raises the sampling weight of medium-accuracy samples and cuts easy and hard ones; and Test-Time Resolution Scaling, which resizes the short side of test images to a fixed length and, in ensemble mode, selects the least-overlapping prediction among three scales.

What would settle it

On a labeled out-of-domain set, run TACO's three-scale ensemble and compare the least-overlap selection against single-scale inference at 672px, oracle selection of the scale with highest IoU, and majority vote. If the least-overlap rule is not substantially better than majority vote and is much worse than oracle, the TTME component is not selecting correctness; the same comparison on LISA and RefGTA would settle whether the 75.1% and 78.7% figures are attributable to the selection rule or to the multi-scale computation itself.

Watch

Extended reading notes

Core claim

The central discovery is that the instability and poor generalization of R1-style reinforcement learning in vision-language models can be traced to a mismatch between the chain of thought and the final answer, and that making the reward a joint function of thought, answer, and ground truth removes the incentive to produce either a long irrelevant chain or a correct-looking answer without reasoning. For grounding tasks this joint reward is realized as a three-way intersection-over-union among the box mentioned in the thinking text, the box in the final answer, and the ground-truth box; for open-ended VQA it is an external supervisor score of the thought-answer pair against the reference. The paper further argues that long-chain exploration collapses not from length itself but from temporary 'dirty' samples whose policy-reference KL divergence spikes, and that masking their gradients while lowering their resampling probability keeps training stable. A test-time multi-scale ensemble that picks the prediction with least overlap across scales is presented as bridging the resolution gap between compressed training images and varied test images.

Load-bearing premise

The headline out-of-distribution gains rest on the assumption that, at test time, the model's prediction with the least overlap across image scales is the correct one; without ground truth to verify that rule, the large LISA and RefGTA improvements shrink to the smaller single-scale numbers.

Editorial extensions

If this is right

  • An R1-style GRPO run with TACO is stable for 1,000 steps on a 3-billion-parameter vision-language model without the response-length collapse seen in the baseline.
  • The gains transfer out of domain: the reported LISA and RefGTA accuracy improvements are larger than the in-domain RefCOCO/+/g gains, suggesting the consistency reward teaches a generalizable answer-grounding habit rather than dataset memorization.
  • Data efficiency improves: training uses only about 1.875% of the available region descriptions, and the adaptive schedule concentrates repeated learning on moderate-difficulty samples.
  • The same single model improves a broad set of VQA benchmarks, including math, charts, diagrams, and OCR, while keeping document VQA essentially flat, indicating the method does not sacrifice general visual QA for grounding.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the three-way IoU is the active ingredient, the recipe should transfer to any task whose output can be decoded into two comparable structured predictions, such as segmentation masks or keypoint sets, where the same thought-versus-answer-versus-ground-truth overlap can be computed.
  • Editorial inference: the least-overlap selection rule in Test-Time Multi-Scale Ensemble is a strong prior that can be tested directly against an oracle that sees ground truth; if the oracle disagrees frequently, a learned scale aggregator or simple majority vote would be a safer default.
  • Editorial inference: the KL-threshold dirty-sample detector suggests that long-chain collapse in RL may be a policy-drift phenomenon rather than a reward-hacking one, which would make the same rollback principle applicable to other RL-trained sequence models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes TACO, an RL fine-tuning method for LVLMs built on GRPO. It introduces a Think-Answer Consistency reward that couples the model's reasoning trace with its final answer and ground truth; a Rollback Resample Strategy that masks high-KL 'dirty' samples and down-weights their sampling rate; an Adaptive Difficulty Sampling schedule that resamples moderate-difficulty examples more often; and a Test-Time Resolution Scaling scheme plus a multi-scale ensemble (TTME) to bridge training/test resolution gaps. Using Qwen2.5-VL-3B, the authors report gains over VLM-R1 and the base model on REC benchmarks (RefCOCO/+/g, LISA, RefGTA) and several VQA benchmarks, with the largest out-of-domain gains attributed to TTME.

Significance. If the results hold, the paper offers a useful recipe for stabilizing long-chain RL in LVLMs and improving data efficiency; the component-level ablations and training-step curves are valuable. The paper is honest about its limitations and does not oversell the ID gains. However, the central empirical claims are currently weakened by an underspecified and internally inconsistent TTME description, a lack of seed-level variance, and unvalidated hand-set hyperparameters. The non-TTME results still show positive gains over VLM-R1, so the core direction is promising but needs substantial revision before the headline numbers can be trusted.

major comments (3)
  1. [3.5, Tables 1, 3, 7] Section 3.5 describes TTME as 'calculates β accuracy rewards through the model' and 'selects the answer with the least number of intersections,' with 'least IoU overlap' for REC. At inference there is no ground truth, so an 'accuracy reward' cannot be computed, and IoU requires an anchor box; the selection rule is therefore not implementable as written. This is load-bearing because the headline OOD numbers (75.1 on LISA, 78.7 on RefGTA in Table 3) and the largest ablation gain in Table 7 (66.5 to 75.1) depend on TTME. In addition, Table 1 reports 'w/ TTME' as 70.81 on LISA, while Table 3 reports 75.1 for 'Ours(w TTME)' and Table 7's final row gives 75.1; these numbers are mutually inconsistent. Please specify the exact reference-free selection rule, justify why the least-overlap prediction is the correct one, clarify whether the ablated 'TTRS' includes TTME or only single-scale TTRS, and resolve the table conflict.
  2. [4.2, Tables 2-7] The paper reports no random seeds, no variances, and no error bars; Table 2's caption says 'The best performance is reported here for each method.' Since GRPO training is stochastic and the claimed margins are small (e.g., +1.7% average over VLM-R1 in Table 2, +5.34 on LISA at 800 steps in Table 4), the results could be within run-to-run noise. Please provide means and standard deviations over at least three independent runs for the main comparisons, or release the exact training scripts and checkpoints so the results can be reproduced.
  3. [3.3-3.4, Eqs. (5)-(7)] All of the new mechanisms rely on hand-set thresholds and coefficients (κ=0.5, γ=0.8, θH=0.5, θL=0.2, αeasy=0.1, αhard=0.8, αmoderate=1.5) with no sensitivity analysis. The KL-based dirty-sample criterion in Eq. (5) is particularly fragile: the KL magnitude depends on training dynamics, and because RRS masks these samples, the reported KL stabilization in Figure 3 is partly by construction. Please include sensitivity sweeps for these hyperparameters and an evaluation of RRS that reports accuracy and KL on samples that are not masked, so the effect of the gradient mask is separated from the effect of the sampling-rate update.
minor comments (5)
  1. [3.4] The offline curation step says samples are 'randomly combined in a 1:2 ratio' but does not state the final training-set size or whether the 1:2 ratio is difficult-to-simple; please clarify.
  2. [3.2, Eq. (4)] The supervisor prompt for VQA mentions 'given corpus' while the equation inputs are Q, T, and GT; please align the prompt text with the notation and explain what the supervisor actually receives.
  3. [Table 4] The SFT comparison lacks training details (epochs, learning rate, data mixture); without these, the SFT-vs-RL comparison is difficult to interpret.
  4. [Figure 2] The caption ends with the incomplete sentence 'reasoning time remains nearly.'; please complete it.
  5. [References] Reference [36] is cited for the 'R1-Vision collection' but the entry names R1-Onevision; please clarify whether these are the same resource or different datasets.

Circularity Check

1 steps flagged · score 2.0 of 10

No central circularity: benchmark gains are external; one minor self-confirmation in RRS's KL-based dirty-sample masking.

  1. other [Section 3.3 (Eq. 5, Eq. 6) and Figure 3(c)]
    "For the “dirty sample” id, RRS applies two measures: Gradient Masking: The gradient of the “dirty sample” id generated by the policy πθ is masked from gradient computations during the current training step."

    Dirty samples are defined by Eq. (5) as those with D_KL(πθ(·|x) ∥ πref(·|x)) > κ. RRS then masks exactly these samples' gradients, so the KL stabilization shown in Figure 3(c) is partly enforced by construction rather than an independent empirical finding. This is a minor internal self-confirmation; it does not manufacture the external benchmark numbers in Tables 2–6, which are measured on held-out test sets and remain independent of the masking rule.

full rationale

The central claims of the paper are empirical: TACO is trained on RefCOCO/+/g and evaluated on held-out RefCOCO/+/g, RefGTA, LISA, and VQA benchmarks. The reported test numbers are measured against external ground truth and are not computed from the training rewards by construction. The TAC reward, RRS, ADS, and TTRS are training/inference mechanisms whose effectiveness is judged by held-out accuracy. The only mild circular flavor is RRS's KL-based dirty-sample masking, which makes part of the KL-stability evidence definitional, and ADS's use of an accuracy reward for sampling that is also related to reported accuracy; neither determines the external benchmark outcomes. The TTME selection rule is underspecified and Table 1/Table 3 are internally inconsistent, but those are correctness and reproducibility concerns, not circularity. There is no load-bearing self-citation chain or imported uniqueness theorem, so the paper is largely self-contained as an empirical comparison.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim rests on several assumptions not proven in the paper: the GRPO objective transfers to multimodal policies, the external VQA supervisor scores consistency reliably, the triple-IoU reward faithfully captures reasoning, and the KL threshold identifies samples that should be masked. Many numerical choices (kappa, gamma, theta, alpha, beta, Starget) are hand-set. None of these are derived, but they are standard engineering choices rather than scientific postulates.

free parameters (8)
  • kappa (dirty-sample KL threshold) = 0.5
    Section 3.3, Eq. 5: used to label samples as dirty and mask gradients; no sensitivity analysis.
  • gamma (dirty resample down-weight) = 0.8
    Section 3.3, Eq. 6: multiplies sampling rate of dirty samples; no sensitivity analysis.
  • theta_H and theta_L (easy or hard accuracy thresholds) = 0.5 and 0.2
    Section 3.4, Eq. 7: classify samples into easy, moderate, and hard; no sensitivity analysis.
  • alpha_easy, alpha_hard, alpha_moderate (ADS sampling multipliers) = 0.1, 0.8, 1.5
    Section 3.4, Eq. 7: adjust sampling rates; no sensitivity analysis.
  • beta (number of test scales) = 3
    Section 3.5: multi-scale ensemble size; chosen without analysis of scale count.
  • Starget (test short-side target length) = 672 pixels
    Section 3.5, Eq. 8: single-scale TTRS target; among 560, 672, and 800 tested, 672 is used in the main results.
  • Offline curation difficulty split ratio = 1:2 difficult to simple
    Section 3.4: base model accuracy used to split 320k samples into 44k difficult and 276k simple, then recombined at a chosen ratio.
  • VQA training set size and step count = 9600 samples, 800 steps
    Section 4.1: sampled from R1-Vision collection; step count aligned with training schedule without formal justification.
assumptions (6)
  • domain assumption GRPO-style policy optimization without a value function is a valid and stable update for LVLMs.
    Preliminary Section 3.1 uses the GRPO objective (Eq. 1) without evidence that it remains stable for multimodal policies; the paper's RRS is introduced precisely because instability occurs.
  • domain assumption The external supervisor Qwen2.5-VL-32B gives reliable semantic-consistency scores in [0,1] for VQA.
    Section 3.2, Eq. 4 uses S(Q,T,GT) as the TAC reward for VQA; the supervisor's reliability is not quantified.
  • domain assumption Three-way IoU between thinking box, answer box, and GT box is a faithful measure of think-answer consistency.
    Section 3.2, Eq. 3 defines R_TAC as triple IoU; assumes the model's emitted think box is a causally meaningful part of reasoning.
  • ad hoc to paper High KL divergence between current and reference policy identifies temporary dirty samples that should be gradient-masked.
    Section 3.3, Eq. 5: the detection rule is a design choice specific to this paper, with kappa manually set.
  • ad hoc to paper Multi-scale predictions concentrate the correct answer in a few scales, and selecting the minimum-overlap prediction is valid.
    Section 3.5 and Table 1: post-hoc observation on LISA; the selection rule is not derived.
  • domain assumption Reweighting samples by current accuracy reward (ADS) improves learning without biasing the final policy.
    Section 3.4: assumes the curriculum does not distort the learned distribution; no theoretical guarantee.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TACO: Think-Answer Consistency for Optimized Long-Chain Reasoning and Efficient Data Learning via Reinforcement Learning in LVLMs." pith.science (2026). https://pith.science/paper/Z6FP4LKY

@misc{pith2026250520777,
  author       = {Pith},
  title        = {Pith review of: TACO: Think-Answer Consistency for Optimized Long-Chain Reasoning and Efficient Data Learning via Reinforcement Learning in LVLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z6FP4LKY}},
  note         = {Machine review of arXiv:2505.20777}
}
read the original abstract

DeepSeek R1 has significantly advanced complex reasoning for large language models (LLMs). While recent methods have attempted to replicate R1's reasoning capabilities in multimodal settings, they face limitations, including inconsistencies between reasoning and final answers, model instability and crashes during long-chain exploration, and low data learning efficiency. To address these challenges, we propose TACO, a novel reinforcement learning algorithm for visual reasoning. Building on Generalized Reinforcement Policy Optimization (GRPO), TACO introduces Think-Answer Consistency, which tightly couples reasoning with answer consistency to ensure answers are grounded in thoughtful reasoning. We also introduce the Rollback Resample Strategy, which adaptively removes problematic samples and reintroduces them to the sampler, enabling stable long-chain exploration and future learning opportunities. Additionally, TACO employs an adaptive learning schedule that focuses on moderate difficulty samples to optimize data efficiency. Furthermore, we propose the Test-Time-Resolution-Scaling scheme to address performance degradation due to varying resolutions during reasoning while balancing computational overhead. Extensive experiments on in-distribution and out-of-distribution benchmarks for REC and VQA tasks show that fine-tuning LVLMs leads to significant performance improvements.

Figures

Figures reproduced from arXiv: 2505.20777 by the authors.

Figure 1
Figure 1. An example of a REC task: Illustrating (center) an enhanced GRPO-based learning loop [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. During training, the TAC reward ensures consistent Think-Answer output. Samples are [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Effectiveness of the Think-Answer Consistency (TAC) reward, comparing TACO, VLM-R1, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. LenGuard-GPC: Length Guarding with Guided-Prompt Consistency for Spatial Reasoning Reinforce Learning

    cs.AI 2026-07 conditional novelty 5.0 of 10

    LenGuard-GPC adds a token-level KL consistency reward between standard and guided prompts, plus a staged length bonus, to GRPO training of Qwen3-VL-8B and reports better accuracy with shorter responses on multi-view s...

Reference graph

Works this paper leans on

52 extracted references · 12 canonical work pages · cited by 1 Pith paper

  1. [1]

    Vision-language models for vision tasks: A survey, 2024

    Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey, 2024

    2024

  2. [2]

    Show, attend and tell: Neural image caption generation with visual attention, 2016

    Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention, 2016

    work page 2016

  3. [3]

    Lawrence Zitnick, Dhruv Batra, and Devi Parikh

    Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2016

    2016

  4. [4]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

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

    arXiv 2023

  5. [5]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report.arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  6. [6]

    Deepseek- vl: Towards real-world vision-language understanding, 2024

    Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, and Chong Ruan. Deepseek- vl: Towards real-world vision-language understanding, 2024

    2024

  7. [7]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024

    2024

  8. [8]

    3d-gpt: Procedural 3d modeling with large language models.arXiv preprint arXiv:2310.12945, 2023

    Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3d-gpt: Procedural 3d modeling with large language models.arXiv preprint arXiv:2310.12945, 2023

    arXiv 2023

Show all 52 references
  1. [9]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    2023

  2. [10]

    Instructblip: Towards general-purpose vision-language models with instruction tuning, 2023

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning, 2023

    2023

  3. [11]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023

    2023

  4. [12]

    Detecting and mitigating hallucination in large vision language models via fine-grained ai feedback

    Wenyi Xiao, Ziwei Huang, Leilei Gan, Wanggui He, Haoyuan Li, Zhelun Yu, Fangxun Shu, Hao Jiang, and Linchao Zhu. Detecting and mitigating hallucination in large vision language models via fine-grained ai feedback. InProceedings of the AAAI Conference on Artificial Intelligence...

    2025

  5. [13]

    Rlthf: Targeted human feedback for llm alignment, 2025

    Yifei Xu, Tusher Chakraborty, Emre Kıcıman, Bibek Aryal, Eduardo Rodrigues, Srinagesh Sharma, Roberto Estevao, Maria Angels de Luis Balaguer, Jessica Wolk, Rafael Padilha, Leonardo Nunes, Shobana Balakrishnan, Songwu Lu, and Ranveer Chandra. Rlthf: Targeted human feedback for ...

    2025

  6. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

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

    2025 arXiv

  7. [15]

    Grpo-lead: A difficulty-aware reinforcement learning approach for concise mathematical reasoning in language models, 2025

    Jixiao Zhang and Chunsheng Zuo. Grpo-lead: A difficulty-aware reinforcement learning approach for concise mathematical reasoning in language models, 2025

    2025

  8. [16]

    Swe-rl: Advancing llm reasoning via reinforcement learning on open software evolution.arXiv preprint arXiv:2502.18449, 2025

    Yuxiang Wei, Olivier Duchenne, Jade Copet, Quentin Carbonneaux, Lingming Zhang, Daniel Fried, Gabriel Synnaeve, Rishabh Singh, and Sida I Wang. Swe-rl: Advancing llm reasoning via reinforcement learning on open software evolution.arXiv preprint arXiv:2502.18449, 2025

    2025 arXiv

  9. [17]

    Towards visual grounding: A survey, 2024

    Linhui Xiao, Xiaoshan Yang, Xiangyuan Lan, Yaowei Wang, and Changsheng Xu. Towards visual grounding: A survey, 2024

    2024

  10. [18]

    Vlm-r1: A stable and generalizable r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, et al. Vlm-r1: A stable and generalizable r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025

    2025 arXiv

  11. [19]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

    2021

  12. [20]

    A versatile vision- language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966, 2023

    J Bai, S Bai, S Yang, S Wang, S Tan, P Wang, J Lin, C Zhou, and J Qwen-VL Zhou. A versatile vision- language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966, 2023. 10

    2023 arXiv

  13. [21]

    Introducing openai o1-preview

    OpenAI. Introducing openai o1-preview. Technical report, OpenAI, 2024. Accessed: 2025-05-03

    2024

  14. [22]

    R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization.arXiv preprint arXiv:2503.10615, 2025

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization.arXiv preprint arXiv:2503.10615, 2025

    2025 arXiv

  15. [23]

    R1-v: Reinforcing super generalization ability in vision-language models with less than $3

    Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. R1-v: Reinforcing super generalization ability in vision-language models with less than $3. Technical report, GitHub, 2025. Accessed: 2025-02-02

    2025

  16. [24]

    Visual-rft: Visual reinforcement fine-tuning.arXiv preprint arXiv:2503.01785, 2025

    Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning.arXiv preprint arXiv:2503.01785, 2025

    2025 arXiv

  17. [25]

    aha moment

    Hengguang Zhou, Xirui Li, Ruochen Wang, Minhao Cheng, Tianyi Zhou, and Cho-Jui Hsieh. R1-zero’s" aha moment" in visual reasoning on a 2b non-sft model.arXiv preprint arXiv:2503.05132, 2025

    2025 arXiv

  18. [26]

    Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning, 2025

    Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Tiancheng Han, Botian Shi, Wenhai Wang, Junjun He, Kaipeng Zhang, Ping Luo, Yu Qiao, Qiaosheng Zhang, and Wenqi Shao. Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based rei...

    2025

  19. [27]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  20. [28]

    Referring expression comprehension: A survey of methods and datasets.IEEE Transactions on Multimedia, 23:4426–4440, 2020

    Yanyuan Qiao, Chaorui Deng, and Qi Wu. Referring expression comprehension: A survey of methods and datasets.IEEE Transactions on Multimedia, 23:4426–4440, 2020

    2020

  21. [29]

    Grec: Generalized referring expression comprehension.arXiv preprint arXiv:2308.16182, 2023

    Shuting He, Henghui Ding, Chang Liu, and Xudong Jiang. Grec: Generalized referring expression comprehension.arXiv preprint arXiv:2308.16182, 2023

    2023 arXiv

  22. [30]

    Vqa: Visual question answering

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. InProceedings of the IEEE international conference on computer vision, pages 2425–2433, 2015

    2015

  23. [31]

    Visual question answering using deep learning: A survey and performance analysis

    Yash Srivastava, Vaishnav Murali, Shiv Ram Dubey, and Snehasis Mukherjee. Visual question answering using deep learning: A survey and performance analysis. InComputer Vision and Image Processing: 5th International Conference, CVIP 2020, Prayagraj, India, December 4-6, 2020, Re...

    2020

  24. [32]

    Generation and comprehension of unambiguous object descriptions

    Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 11–20, 2016

    2016

  25. [33]

    Modeling context in referring expressions

    Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. InComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69–85. Springer, 2016

    2016

  26. [34]

    Generating easy-to-understand referring expressions for target identifications

    Mikihiro Tanaka, Takayuki Itamochi, Kenichi Narioka, Ikuro Sato, Yoshitaka Ushiku, and Tatsuya Harada. Generating easy-to-understand referring expressions for target identifications. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5794–5803, 2019

    2019

  27. [35]

    Lisa: Reasoning segmentation via large language model

    Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579–9589, 2024

    2024

  28. [36]

    R1-onevision: A unified benchmark for vision-language reasoning and generation, June

    Zhelun Shen, Zitian Chen, Yushi Liu, Meiqi Chen, Zongyi Liu, Runlian Shen, Leilei Sun, Haozhe Zhao, Hengfei Wang, Yuxiang Wei, Junchi Yan, Hongyan Liu, Xiaodan Liang, Ming-Hsuan Yang, and Anton van den Hengel. R1-onevision: A unified benchmark for vision-language reasoning and...

  29. [37]

    Are we on the right way for evaluating large vision-language models?arXiv preprint arXiv:2403.20330, 2024

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models?arXiv preprint arXiv:2403.20330, 2024

    2024 arXiv

  30. [38]

    A diagram is worth a dozen images

    Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. InEuropean conference on computer vision, pages 235–251. Springer, 2016

    2016

  31. [39]

    Infographicvqa

    Minesh Mathew, Viraj Bagal, Rubèn Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697–1706, 2022

    2022

  32. [40]

    Towards vqa models that can read

    Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317–8326, 2019

    2019

  33. [41]

    Docvqa: A dataset for vqa on document images

    Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200–2209, 2021. 11

    2021

  34. [42]

    Measuring multimodal mathematical reasoning with math-vision dataset.Advances in Neural Information Processing Systems, 37:95095–95169, 2024

    Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hong- sheng Li. Measuring multimodal mathematical reasoning with math-vision dataset.Advances in Neural Information Processing Systems, 37:95095–95169, 2024

    2024

  35. [43]

    Mmbench: Is your multi-modal model an all-around player?, 2024

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, Kai Chen, and Dahua Lin. Mmbench: Is your multi-modal model an all-around player?, 2024

    2024

  36. [44]

    On the hidden mystery of ocr in large multimodal models.arXiv e-prints, pages arXiv–2305, 2023

    Yuliang Liu, Zhang Li, Biao Yang, Chunyuan Li, Xucheng Yin, Cheng-lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery of ocr in large multimodal models.arXiv e-prints, pages arXiv–2305, 2023

    2023

  37. [45]

    Hierarchical alignment- enhanced adaptive grounding network for generalized referring expression comprehension.arXiv preprint arXiv:2501.01416, 2025

    Yaxian Wang, Henghui Ding, Shuting He, Xudong Jiang, Bifan Wei, and Jun Liu. Hierarchical alignment- enhanced adaptive grounding network for generalized referring expression comprehension.arXiv preprint arXiv:2501.01416, 2025

    2025 arXiv

  38. [46]

    Internvl2: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024

    OpenGVLab Team. Internvl2: Better than the best—expanding performance boundaries of open-source multimodal models with the progressive scaling strategy, 2024

    2024

  39. [47]

    Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer.arXiv preprint arXiv:2408.03326, 2024

    2024 arXiv

  40. [48]

    Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, et al. Mm1. 5: Methods, analysis & insights from multimodal llm fine-tuning.arXiv preprint arXiv:2409.20566, 2024

    2024 arXiv

  41. [49]

    Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models.arXiv preprint arXiv:2409.17146, 2024

    Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models.arXiv preprint arXiv:2409.17146, 2024

    2024 arXiv

  42. [50]

    Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024

    Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800, 2024

    2024 arXiv

  43. [51]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024. 12

    2024 arXiv

  44. [2024]

    Code and data available athttps://github.com/Fancy-MLLM/R1-Onevision

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.