Pith. sign in

REVIEW 4 major objections 5 minor 42 references

A single image-editing operator can make a vision-language model change its answer in some samples and not others; SIVA-RL uses the observed reward drop, not the operator type, to decide whether each sample should train sensitivity or invar

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 03:16 UTC pith:SHTWWYA2

load-bearing objection A real, well-scoped contribution with honest caveats, but single-run numbers and an untested routing signal keep the mechanism plausible rather than proven. the 4 major comments →

arxiv 2607.13931 v1 pith:SHTWWYA2 submitted 2026-07-15 cs.CV

SIVA-RL: Sensitivity-Invariance Visual Alignment for Multimodal Reinforcement Learning

Cheng Tang , Junzhi Ning , Min Cen , Wei Li , Xinyi Zeng , Pinxian Zeng , Rongbin Li , Qiming Zhu
show 4 more authors
Yuqiang Li Junjun He Yirong Chen Ming Hu
This is my paper
classification cs.CV
keywords multimodal reinforcement learningvisual groundingoutcome-conditioned supervisionPatchSwapsensitivity-invariance alignmentRLVRreward-drop routingvision-language model
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 argues that when reinforcement learning rewards only final answers, a vision-language model can score well without actually looking at the image. Visual-intervention methods try to fix this by comparing behavior on original and modified images, but they assume the type of intervention determines what the model should learn. SIVA-RL shows this assumption fails: the same patch-swapping operation causes large reward drops in some samples and no change in others. Instead of labeling pairs by operator, SIVA-RL labels them by the measured drop—large drops train sensitivity to visual evidence, small drops train clean-anchored invariance, and ambiguous pairs are down-weighted. On nine multimodal reasoning benchmarks, this outcome-conditioned routing improves both 3B and 7B models over their matched RL baselines in every configuration, with the largest gains on vision-dependent tasks.

Core claim

The central discovery is that intervention-defined supervision in multimodal RL is ambiguous: the same visual operator (black-image, random-mask, or PatchSwap) produces both answer-drop and answer-stable outcomes within the clean-correct subset of a benchmark. SIVA-RL therefore separates intervention construction from supervision assignment. It builds localized within-image PatchSwap interventions, uses a frozen audit policy to compute the clean-to-intervention reward drop for each sample, and converts that drop into soft routing weights: large drops drive a margin-bounded sensitivity objective, low drops drive a clean-anchored invariance objective, and ambiguous pairs are down-weighted. Und

What carries the argument

The central mechanism is outcome-conditioned soft routing driven by the clean-to-intervention reward drop Δ, computed by a frozen audit policy over paired clean and PatchSwap views. PatchSwap is a localized within-image operation that replaces selected cells with distant same-image patches, preserving global style while disrupting local evidence. The reward drop is mapped through ramp functions into weights that respectively push the clean-intervention token-level divergence above a sensitivity margin or pull it below an invariance margin, with an abstention band between thresholds. Re-scoring the already-sampled clean responses under both visual conditions makes the auxiliary loss cheap, an

Load-bearing premise

The load-bearing premise is that a large clean-to-intervention reward drop, measured by a frozen audit policy, faithfully marks sensitivity to the swapped visual content, and a small drop marks insensitivity—but the paper itself notes a big drop can come from intervention artifacts and a small drop can come from language-prior shortcuts, so if these confounds dominate, the routing would train the policy to attend to the wrong cues.

What would settle it

Take a benchmark where many correct answers are derivable from the question text alone. Train with SIVA-RL, then probe the resulting policy with the image blanked or swapped in regions known to be irrelevant to the answer. If accuracy stays high on the text-only probe after training (i.e., the policy still relies on language priors), or if the policy's correctness degrades when an irrelevant region is removed (i.e., sensitivity was attached to artifacts), the routing signal is not doing what the paper claims. More directly: compare the set of sensitivity-routed samples to a human-labeled map o

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

If this is right

  • Visual-intervention RL methods should route supervision by the observed sample-level effect of an intervention, not by the operator's type.
  • The same PatchSwap operator can serve both sensitivity and invariance training simultaneously, so intervention design can focus on supplying informative pairs rather than on prescribing an objective.
  • SIVA-RL's gains concentrate on vision-dependent benchmarks (up to +8.79 points), indicating that outcome-conditioned alignment specifically improves use of visual evidence rather than giving a generic RL boost.
  • Because SIVA-RL is a plug-in auxiliary objective on top of GRPO or DAPO, it can be layered onto future RLVR backbones without changing the base algorithm.
  • The approach's robustness to construction hyperparameters (a wide plateau around optimal settings) suggests it can be deployed without costly per-dataset tuning.

Where Pith is reading between the lines

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

  • The reward-drop routing signal could be replaced or augmented with token-level attribution or attention-based grounding measures, which might better distinguish artifact-induced drops from genuine reliance on visual evidence.
  • The same outcome-conditioned principle should transfer to other perturbation families (counterfactual edits, text interventions) and other modalities (audio, video), where operator heterogeneity is likely just as severe.
  • A stronger test of the paper's premise would be to verify that sensitivity-routed pairs actually correspond to human-annotated task-relevant image regions; if routing agrees with human evidence maps, the method is a step toward certificated visual grounding rather than just better benchmark scores.
  • One could test the limits of the invariance branch by constructing adversarial low-drop pairs where the model's answer is stable only because it ignores the image entirely (language prior); the current clean-correctness gate does not fully rule this out.

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

4 major / 5 minor

Summary. The paper proposes SIVA-RL, a method for multimodal reinforcement learning that augments a base GRPO or DAPO objective with an auxiliary loss derived from paired clean/intervened image rollouts. The key novelty is to replace operator-conditioned supervision (e.g., all corrupted images should produce invariance) with sample-wise routing based on the observed clean-to-intervention reward drop computed by a frozen audit policy. PatchSwap constructs localized within-image interventions; a clean-correctness gate and soft thresholds then route each pair to a sensitivity loss (for large drops) or a clean-anchored invariance loss (for low drops), with an abstention band for ambiguous pairs. Experiments on nine benchmarks with Qwen2.5-VL-3B/7B and GRPO/DAPO backbones report consistent improvements over matched baselines, with larger gains on vision-dependent subsets, plus ablations and training diagnostics.

Significance. If the central claim is established, SIVA-RL would be a useful contribution to multimodal RL: it draws attention to a real limitation of answer-level rewards, proposes a concrete sample-wise treatment of visual interventions, and provides a framework that is agnostic to the backbone RL algorithm. The paper also includes a genuinely informative audit showing that identical interventions produce heterogeneous outcome transitions, and the idea of separating intervention construction from supervision assignment is conceptually sound. However, the empirical validation as presented does not yet confirm the load-bearing assumption that the reward-drop routing signal tracks visual evidence rather than artifacts or language priors; the current evidence could also be explained by generic auxiliary-loss benefits. The paper is transparent about several of these limitations in the text, which is to its credit, but the missing control experiments are central to the contribution.

major comments (4)
  1. [§4.1, Table 1 and Table 7] The main claim that SIVA-RL improves 'every setting' rests on single-run Avg@8 numbers with no variance, confidence intervals, or multiple seeds for any configuration. In addition, the PatchSwap hyperparameters η=0.6 and d_min=0.25 are selected in Table 7 using the same benchmark suite (GRPO-3B overall Avg@8), and these same values are then used in the main results of Table 1. This is test-set selection without a validation split. The authors should report seed variance and either use a held-out validation set for hyperparameter choice or demonstrate that the main conclusions are insensitive to the chosen values within the flat regions they claim.
  2. [§3.3, Eqs. (8)–(10)] The central mechanism is the assumption that the reward drop Δ_i is a reliable per-sample routing signal for visual evidence dependence. The paper itself explicitly acknowledges (Section 1) that a large drop may be due to intervention-induced artifacts and that a stable outcome may reflect language-prior reliance. Yet no control experiment tests whether the routing signal carries the claimed information. The authors should add at least random routing, inverted routing, and, ideally, an oracle-routing control: if random or inverted routing yields comparable gains on the same auxiliary losses, the outcome-conditioned routing is not validated. Without such controls, the observed improvements could come from the auxiliary sensitivity/invariance losses themselves rather than from their sample-wise assignment.
  3. [§3.3, Eq. (7) and Table 6] The routing weights are computed from M=4 sampled audit rollouts per view, and Section 3.3 states no stability analysis. With only four samples, Δ_i is a noisy estimator of the policy's expected clean-to-intervention reward change; this noise is then thresholded by intervals as narrow as [0.0, 0.05] for invariance routing. The paper reports no audit-seed variance, no M larger than 4, and no sensitivity of routing rates to the audit temperature or top-p. A simple experiment varying M (e.g., M=4 vs M=16 vs M=64) or repeating the same audit with different seeds would clarify whether the downstream training signal is stable. This is load-bearing because the soft routing weights directly determine the direction and strength of L_sen and L_inv.
  4. [§3.4, Eq. (8) and (13)] The clean-correctness gate b_i only checks that R_cl ≥ τ_c, so pairs that are correct through a language prior are not excluded. On such pairs, the invariance loss L_inv actively reduces D_{i,g}, which can entrench a visually ungrounded response. The paper acknowledges that the gate does not certify visual-evidence use (Section 1, Section 3.4), but it does not provide an analysis or experiment to show this failure mode is rare or that the invariance branch is not dominated by language-prior-stable pairs. The authors should either add a control that measures the fraction of routed low-drop pairs attributable to language priors (e.g., black-image or text-only probes) or temper the claim that invariance alignment improves visual grounding rather than merely consistency.
minor comments (5)
  1. [§3.2, Eq. (5)] The definition of S_i(u) requires C_{i,v}=0, but when η=1.0 the candidate set is the whole grid and the source set is empty; the fallback rule is only described in Appendix A. This is fine, but the main text should point to the fallback rule earlier, since Figure 4 discusses η=1.0 as a stress test.
  2. [§4.1, Table 1] The notation 'Avg@8' is used without definition in the main text; it is only clarified later in Table 5's validation decoding multiplicity. Please define it at first use, and state whether the reported numbers are averages over 8 sampled rollouts per prompt.
  3. [§3.4, Eq. (12)] The quantity d_{i,g,t} is called a 'Monte Carlo surrogate' for token-level KL divergence, but the formula is the Bregman divergence associated with the exponential family; the relation to a Monte Carlo estimate is not obvious. A short derivation or citation would help.
  4. [§4.3, Table 4] The stepwise ablation labels the final row as 'SIVA-RL' after adding invariance; this is clear, but the reader would benefit from a column showing the exact loss coefficients used in the 'Soft Sens. + Warmup (w/o Inv.)' row versus the full method, since the incremental gain of +2.37% is associated with both invariance and any re-weighting changes.
  5. [Appendix D, Table 8] The medical-domain comparison is against the pretrained base model rather than a matched RL baseline, and the authors explicitly caveat this in Section D.2. This is appropriate, but the main text should not imply that these results validate SIVA-RL's grounding mechanism; the caveat should also appear where medical results are summarized.

Circularity Check

1 steps flagged

Main benchmark gains are empirical and self-contained; only a training-dynamics diagnostic is self-confirmatory by construction.

specific steps
  1. other [Appendix C.2, Routing-Branch Separation (and Section 3.4, Eqs. 12-13)]
    "The separation confirms the mechanistic claim in Section 4.4: the same PatchSwap operator keeps feeding both answer-changing and low-drop pairs rather than degenerating into a uniform perturbation loss."

    The separation is the direct effect of the two margin losses: L_sen pushes D up for high-Delta pairs and L_inv pushes D down for low-Delta pairs (Eq. 13). Observing that sensitivity-routed divergence is higher than invariance-routed divergence is therefore observing the loss's own operation, not independent evidence that Delta tracks visual evidence. The paper itself notes (Sec. 1) that large drops may arise from intervention-induced artifacts and stable outcomes from language priors, so this diagnostic cannot certify grounding. It is a sanity check, not a derivation; the central benchmark comparisons remain external and non-circular.

full rationale

SIVA-RL's central claims are measured on external benchmarks (Table 1) against matched RL baselines; no equation predicts those numbers from the method's own inputs. The routing weights (Eq. 10) are calculated from audit-policy reward drops and then used to weight auxiliary alignment terms; this is a feedback loop (routing signal and base RL reward share the same verifier) and a statistical weakness (M=4 rollouts, no routing controls), but not a tautology. No load-bearing self-citation is present: the PAPO perception branch is cited to prior external work and the MM-Eureka comparison is a baseline, not a premise. The paper is unusually candid in Appendix D.2 that medical accuracy alone is not evidence of visual grounding and in Sec. 1 that reward drops may reflect artifacts rather than evidence. The only by-construction element is the Appendix C.2 'confirmation' of route separation, which merely restates the margin objectives' behavior. This does not affect the empirical main result. Overall score 1: no significant circularity, one minor self-confirmatory diagnostic.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard RLVR assumptions plus the heuristic that answer-level reward drop is a reliable routing signal. The many hyperparameters are hand-chosen or tuned on the evaluation benchmarks, and the method introduces no new physical or mathematical entities.

free parameters (8)
  • reward-drop routing thresholds (delta_lo_sen, delta_hi_sen, delta_lo_inv, delta_hi_inv) = 0.05, 0.5, 0.0, 0.05
    Define the sensitivity and invariance bands; chosen by hand, no principled derivation from first principles.
  • sensitivity margin m_sen and invariance margin m_inv = 0.05, 0.01
    Target divergence levels for the two alignment branches; tuned to make the method work.
  • loss coefficients lambda_sen, lambda_inv = 0.005, 0.001
    Balance the auxiliary alignment losses against the base RL objective.
  • clean correctness threshold tau_c = 0.75
    Gate for deciding which clean-intervention pairs are valid; chosen by hand.
  • PatchSwap candidate-cell ratio eta = 0.6
    Fraction of cells eligible for replacement; selected by sweeping on the same benchmark averages (Table 7).
  • PatchSwap replacement probability rho = 0.75
    Probability that an eligible cell becomes an active replacement target; fixed by hand.
  • PatchSwap minimum source distance d_min = 0.25
    Minimum normalized grid distance for source cells; selected by sweeping on the same benchmark averages (Table 7).
  • audit rollouts per view M = 4
    Number of sampled audit decodings per view; affects noise level of the reward-drop estimate.
axioms (5)
  • domain assumption The answer-level verifier reward is a meaningful and sufficient training signal for multimodal reasoning quality.
    Foundational to RLVR; the whole method optimizes answer-level correctness.
  • domain assumption PatchSwap interventions preserve enough global image context to be in-distribution while removing local evidence.
    Section 3.2; the construction relies on this to provide useful training pairs.
  • domain assumption The observed clean-to-intervention reward drop (Delta) is a usable proxy for whether the model's prediction depends on the edited visual evidence.
    Section 3.3; the paper explicitly states this is a routing signal, not a semantic certificate.
  • domain assumption The token-level divergence surrogate D computed from clean-rollout tokens approximates the full-vocabulary KL divergence.
    Section 3.4, Eq. (12); only tokens from the clean rollout are used, so the estimate is biased.
  • domain assumption The frozen audit policy provides a stable paired outcome estimate during training.
    Section 3.3; stopping gradients avoids feedback but still uses an evolving policy snapshot.

pith-pipeline@v1.3.0-alltime-deepseek · 23543 in / 11906 out tokens · 106931 ms · 2026-08-02T03:16:00.145579+00:00 · methodology

0 comments
read the original abstract

Reinforcement learning with verifiable rewards (RLVR) drives multimodal reasoning, but answer-level correctness does not guarantee that a vision-language model grounds its predictions in visual evidence. Existing visual-intervention methods contrast policy behavior on original and modified images, yet assign supervision by the type of intervention rather than its observed effect. This assumption fails: identical operators produce heterogeneous outcomes across samples. We propose SIVA-RL, a Sensitivity-Invariance Visual Alignment framework that replaces operator-conditioned regularization with sample-wise, outcome-conditioned supervision. SIVA-RL constructs localized interventions through token-aligned, distance-constrained within-image PatchSwap. A frozen audit policy then scores each clean-intervention pair, and the observed reward drop becomes soft routing weights. Large-drop pairs drive sensitivity alignment, low-drop pairs drive clean-anchored invariance alignment, and ambiguous pairs are down-weighted. This design decouples intervention construction from supervision assignment and is compatible with both GRPO and DAPO backbones. Across nine multimodal reasoning benchmarks spanning mathematical, logical, and vision-dependent tasks, SIVA-RL improves 3B and 7B models over matched RL baselines in every setting. It yields an 8.79 percentage-point gain on vision-dependent reasoning and up to 14.9% relative overall improvement across all four GRPO- and DAPO-based configurations.

Figures

Figures reproduced from arXiv: 2607.13931 by Cheng Tang, Junjun He, Junzhi Ning, Min Cen, Ming Hu, Pinxian Zeng, Qiming Zhu, Rongbin Li, Wei Li, Xinyi Zeng, Yirong Chen, Yuqiang Li.

Figure 1
Figure 1. Figure 1: Paired intervention audit and SIVA-RL overview. (a) Visual interventions induce four paired-outcome categories. (b) Black-image, random-mask, and PatchSwap interventions all produce mixtures of outcome-drop and outcome-stable cases among the 906 clean-correct examples. (c) SIVA-RL routes each valid pair toward sensitivity or clean-anchored consistency according to its observed outcome change. Our paired in… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SIVA-RL. The upper branch represents the standard clean-image RLVR backbone. (a) Token-aligned, distance-constrained PatchSwap constructs a localized candidate intervention x˜i . (b) A frozen audit policy performs paired decoding and converts the clean-to￾intervention reward drop into sensitivity and invariance weights. (c) The clean rollout responses are re-scored under both visual conditions.… view at source ↗
Figure 3
Figure 3. Figure 3: Training-time outcome-conditioned routing diagnostics. We report the sensitivity rout￾ing rate, invariance routing rate, and mean paired answer-reward drop for the four backbone–scale configurations. The uniform-sensitivity warm-up raises performance to 52.64. Early-stage optimization therefore benefits from a stabilized alignment signal before outcome-conditioned routing. Invariance align￾ment for clean-a… view at source ↗
Figure 4
Figure 4. Figure 4: Hyperparameter sensitivity. Macro Avg@8 under varying η and dmin. Hyperparameter Sensitivity [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Math-domain reward-accuracy dynamics across the two-stage schedule. The dashed vertical line marks the transition from the 200-step PatchSwap/PAPO-style warm-up to the 100-step SIVA-RL stage, with curves shown for GRPO-3B, DAPO-3B, GRPO-7B, and DAPO-7B. C.1 REWARD–ACCURACY DYNAMICS The routing stage reignites a plateauing warm-up [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: intervention divergence dynamics during the math-domain SIVA-RL stage. The two panels report mean divergence D for sensitivity-routed and invariance-routed pairs, respectively. Each panel reports how the re-scored clean responses behave under PatchSwap interventions across backbone and model scales. drop pairs apart. The invariance branch starts higher in the first few steps and is then driven down toward … view at source ↗
Figure 7
Figure 7. Figure 7: Visual recovery case. SIVA-RL corrects a baseline OCR/program-reading error by using the rendered loop, while the black-image response remains incorrect. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Misleading black-image correctness. The baseline’s correct black-image answer is not evidence of robust visual reasoning; it masks a clean-image failure. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: PatchSwap diagnostic case. A localized within-image replacement reveals a baseline stability gap that the black-image probe does not reveal. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Additional PatchSwap case on geometry. The localized intervention uncovers a base￾line error that is absent on the clean image. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Residual failure on dense multi-instance counting. Even after SIVA-RL, the model can over-count similar objects on the clean image and become more unstable under local or global visual perturbations. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_11.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

42 extracted references · 16 linked inside Pith

  1. [1]

    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

  2. [2]

    arXiv preprint arXiv:2501.12948 , year=

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

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2503.06749 , year=

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

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2505.16673 , year=

    R1-sharevl: Incentivizing reasoning capability of multimodal large language models via share-grpo , author=. arXiv preprint arXiv:2505.16673 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2503.07365 , year=

    Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning , author=. arXiv preprint arXiv:2503.07365 , year=

    Pith/arXiv arXiv

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    Vl-rethinker: Incentivizing self-reflection of vision-language models with reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Making the v in vqa matter: Elevating the role of image understanding in visual question answering , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  8. [8]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Don't just assume; look and answer: Overcoming priors for visual question answering , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  9. [9]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Counterfactual vqa: A cause-effect look at language bias , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  10. [10]

    arXiv preprint arXiv:2505.14677 , year=

    Visionary-r1: Mitigating shortcuts in visual reasoning with reinforcement learning , author=. arXiv preprint arXiv:2505.14677 , year=

    arXiv

  11. [11]

    arXiv preprint arXiv:2506.07218 , year=

    Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward , author=. arXiv preprint arXiv:2506.07218 , year=

    arXiv

  12. [12]

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

    R1-VL: Learning to Reason with Multimodal Large Language Models via Step-wise Group Relative Policy Optimization , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  13. [13]

    arXiv preprint arXiv:2507.06448 , year=

    Perception-aware policy optimization for multimodal reasoning , author=. arXiv preprint arXiv:2507.06448 , year=

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2510.09285 , year=

    Spotlight on token perception for multimodal reinforcement learning , author=. arXiv preprint arXiv:2510.09285 , year=

    arXiv

  15. [15]

    arXiv preprint arXiv:2604.09349 , year=

    Visually-Guided Policy Optimization for Multimodal Reasoning , author=. arXiv preprint arXiv:2604.09349 , year=

    Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2606.08708 , year=

    PRPO: Perception-Reinforced Policy Optimization via Token-Level Dynamic Advantage Reshaping , author=. arXiv preprint arXiv:2606.08708 , year=

    Pith/arXiv arXiv

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Noisyrollout: Reinforcing visual reasoning with data augmentation , author=. Advances in Neural Information Processing Systems , volume=

  18. [18]

    arXiv e-prints , pages=

    Vision matters: Simple visual perturbations can boost multimodal math reasoning , author=. arXiv e-prints , pages=

  19. [19]

    arXiv preprint arXiv:2506.17267 , year=

    Cf-vlm: Counterfactual vision-language fine-tuning , author=. arXiv preprint arXiv:2506.17267 , year=

    Pith/arXiv arXiv

  20. [20]

    Inter-gps: Interpretable geometry problem solving with formal language and symbolic reasoning , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages=

  21. [21]

    International Conference on Learning Representations , volume=

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts , author=. International Conference on Learning Representations , volume=

  22. [22]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    We-math: Does your large multimodal model achieve human-like mathematical reasoning? , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  23. [23]

    European Conference on Computer Vision , pages=

    Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  24. [24]

    arXiv preprint arXiv:2407.04973 , year=

    Logicvista: Multimodal llm logical reasoning benchmark in visual contexts , author=. arXiv preprint arXiv:2407.04973 , year=

    Pith/arXiv arXiv

  25. [25]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  26. [26]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  27. [27]

    2024 , month = may, howpublished =

    Hello. 2024 , month = may, howpublished =

    2024

  28. [28]

    2024 , month = jul, howpublished =

    2024

  29. [29]

    2024 , month = dec, howpublished =

    Introducing. 2024 , month = dec, howpublished =

    2024

  30. [30]

    Shuai Bai and Keqin Chen and Xuejing Liu and Jialin Wang and Wenbin Ge and Sibo Song and Kai Dang and Peng Wang and Shijie Wang and Jun Tang and Humen Zhong and Yuanzhi Zhu and Mingkun Yang and Zhaohai Li and Jianqiang Wan and Pengfei Wang and Wei Ding and Zheren Fu and Yiheng Xu and Jiabo Ye and Xi Zhang and Tianbao Xie and Zesen Cheng and Hang Zhang and...

  31. [31]

    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

  32. [32]

    arXiv preprint arXiv:2408.03326 , year=

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

    Pith/arXiv arXiv

  33. [33]

    arXiv preprint arXiv:2509.23661 , year=

    Llava-onevision-1.5: Fully open framework for democratized multimodal training , author=. arXiv preprint arXiv:2509.23661 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2509.00676 , year=

    Llava-critic-r1: Your critic model is secretly a strong policy model , author=. arXiv preprint arXiv:2509.00676 , year=

    Pith/arXiv arXiv

  35. [35]

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

    R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  36. [36]

    arXiv preprint arXiv:2505.12504 , year=

    CPGD: Toward Stable Rule-based Reinforcement Learning for Language Models , author=. arXiv preprint arXiv:2505.12504 , year=

    Pith/arXiv arXiv

  37. [37]

    Training reasoning model with dynamic advantage estimation on reinforcement learning , author=

  38. [38]

    arXiv preprint arXiv:2504.11468 , year=

    Sft or rl? an early investigation into training r1-like reasoning large vision-language models , author=. arXiv preprint arXiv:2504.11468 , year=

    Pith/arXiv arXiv

  39. [39]

    arXiv preprint arXiv:2506.04559 , year=

    Reasoning-Aligned Perception Decoupling for Scalable Multi-modal Reasoning , author=. arXiv preprint arXiv:2506.04559 , year=

    arXiv

  40. [40]

    Advances in Neural Information Processing Systems , volume=

    Sota with less: Mcts-guided sample selection for data-efficient visual reasoning self-improvement , author=. Advances in Neural Information Processing Systems , volume=

  41. [41]

    Advances in Neural Information Processing Systems , volume=

    Dapo: An open-source llm reinforcement learning system at scale , author=. Advances in Neural Information Processing Systems , volume=

  42. [42]

    arXiv preprint arXiv:2601.06801 , year=

    Thinking with Deltas: Incentivizing Reinforcement Learning via Differential Visual Reasoning Policy , author=. arXiv preprint arXiv:2601.06801 , year=

    arXiv