acceptodds
Under review as a conference paper at ICLR 2027

Visual Sensitivity Is Not Claim Retractability: Persistence-Aware Credit Assignment for Multimodal Reinforcement Learning

Abstract

Reinforcement Learning with Verifiable Rewards (RLVR) has been extended to Large Vision–Language Models (LVLMs), and perception-aware methods encourage policies to rely more strongly on visual evidence. Yet stronger visual dependence does not guarantee that individual visual claims are supported by the image: 27.81% of answer-correct base-model trajectories across four benchmarks contain at least one unsupported direct visual claim, which outcome-level RL rewards along with the rest of the trajectory. We introduce a fixed-rollout counterfactual diagnostic that separates Evidence-Function Sensitivity (EFS), how strongly the model responds to an intervened image, from claim persistence, whether it keeps supporting the same claim once its visual evidence is disrupted. The diagnostic reveals Sensitivity–Persistence Decoupling (SPD): under DAPO and VPPO, visual sensitivity increases and aggregate retractability improves, yet unsupported claims become significantly more persistent, whereas GRPO raises sensitivity without this deterioration. We propose Persistence-Aware Credit Gating (PACG), a conservative mechanism that attenuates positive credit for unusually persistent visual claims while leaving negative updates unchanged. It requires no supported/unsupported labels and adds no inference cost. PACG improves unsupported-claim retractability for both optimizers, raises the nine-benchmark average of DAPO from 58.1% to 59.9% and of VPPO from 59.8% to 60.9% on Qwen2.5-VL-7B, transfers to a larger scale and a newer backbone, and the accuracy of HallusionBench also improves consistently. These results suggest that visual dependence and claim retractability are complementary dimensions of multimodal credit assignment.

Primary area:
foundation or frontier models, including LLMs
Keywords:
multimodal large language models, reinforcement learning, multimodal reasoning, credit assignment, visual grounding, visual hallucination

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_WnSxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies visual claims in multimodal reasoning traces, including those that reach correct final answers. It introduces a fixed-rollout counterfactual diagnostic that distinguishes visual sensitivity from claim persistence, showing that the two can diverge under certain RL optimizers. The proposed Persistence-Aware Credit Gating (PACG) uses residual-calibrated likelihood changes under image corruption to attenuate positive credit for persistent visual claims. Experiments across several benchmarks and backbones show performance improvements.

Critical strengths

-The paper addresses an important credit-assignment problem, that a correct final answer does not guarantee that intermediate visual claims are supported.

-PACG is a reasonably simple and conservative modification. It attenuates positive credit on selected visual claims while preserving negative advantages and other tokens.

-The optimizer-dependent findings, including the different behavior of GRPO versus DAPO and VPPO, are informative.

Critical weaknesses

-There are some novelty concerns regarding SRPO. SRPO already uses signed log-likelihood contrasts between original and masked images to modulate perception-token credit. While Table 3 ablated the residual calibration and positive-only gating within PACG, it does not show their incremental benefit over SRPO. A direct comparison using the same backbone and training settings with SRPO would help the reader to identify its contribution and real value. In addition, the overall improvements on all benchmark are quite minor compared to the baseline of DAPO and VPPO.

-The persistence metric measures whether a claim becomes less likely after the image is corrupted, but this does not mean the model would change or stop making that claim in a new response. The corruption may also affect unrelated image regions rather than the evidence for that claim. The existing controls and free-generation audit are helpful, but a stronger test would be that removing the relevant evidence and check whether the model changes its claim.

Questions for authors

See the weaknesses.

3 comments
AuthorsOpenReview ↗

The response to Reviewer WnSx

Dear Reviewer,

I hope this message finds you well. We hope our responses have addressed your concerns. Your further comments are highly valuable for guiding the revision of our manuscript. Should you have any remaining questions or concerns, we welcome further discussion and look forward to your reply.

Thank you for your time and effort in reviewing our paper!

Thank you for the response. I think my main concerns on novelty and persistence metric are solved. I will increase my score.

AuthorsOpenReview ↗

The response to Reviewer WnSx

Thank you very much for your careful evaluation and positive feedback. We greatly appreciate your time and efforts reviewing our manuscript, and we are glad that our revisions have addressed your core concerns regarding novelty and the persistence metric, We wish you all the best!

Reviewer_rTV5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper investigates credit assignment in multimodal RLVR and introduces a fixed-rollout counterfactual diagnostic that separates Evidence-Function Sensitivity (claim-level token KL) from residual-calibrated claim persistence under image intervention. Using fixed Base responses on CrossBench4-2000, the authors report Sensitivity–Persistence Decoupling (SPD), where DAPO and VPPO increase visual sensitivity but worsen the persistence of unsupported claims, unlike GRPO. To mitigate this decoupling, the authors propose Persistence-Aware Credit Gating (PACG), which employs a frozen router to locate direct visual claims and smoothly downweights only positive token credit on persistent claims without requiring support labels. Evaluated across nine multimodal reasoning benchmarks over three seeds on Qwen2.5-VL-7B, PACG improves average accuracy by +1.8 for DAPO and +1.1 for VPPO with additional evaluations on 32B, 2B-Thinking backbones, HallusionBench, and free-generation claim audits.

Critical strengths
  • The counterfactual fixed-rollout diagnostic cleanly separates unsigned visual sensitivity from directional claim persistence via residual-token calibration and question-clustered bootstrap intervals. The empirical finding of Sensitivity–Persistence Decoupling (Figure 4, Table 5) exposes an overlooked failure mode of sequence-level credit assignment in multimodal RLVR, providing a valuable and reusable conceptual framework for the community.
  • Persistence-Aware Credit Gating demonstrates consistent, statistically significant accuracy improvements across nine multimodal reasoning benchmarks over three training seeds for both DAPO and VPPO without incurring inference-time overhead. The performance gains successfully transfer across model scales and distinct model families, demonstrating broad empirical robustness.
Critical weaknesses
  • The paper does not establish that accuracy gains come from persistence gating rather than simply scaling down positive gradients on visual tokens. In Table 4, the default gate keeps an average of 81% of the positive credit on routed visual claims, but no baseline tests whether simply multiplying routed positive advantages by a constant 0.81 achieves the same improvement without any persistence calculation. Moreover, in Table 12, DAPO+PACG yields the largest accuracy gain of +1.8 points while barely changing unsupported persistence by −0.0151 and shifting the image-versus-text DID in the opposite direction by +0.0059. Similarly, in Table 3, removing residual calibration still recovers 59.0% accuracy compared to 59.9% for full PACG, a gap comparable to the ±0.52% seed standard deviation reported in Table 17.
  • Sensitivity–Persistence Decoupling is demonstrated exclusively by teacher-forcing checkpoints on fixed Base rollouts with global image masking (Section 5.4, Appendix C.1), where contradicted claims regressing toward language priors under masking can artificially mimic persistence. Furthermore, the paper's own free-generation audit reveals that unsupported claim rates already decline during RL on each policy's actual generations (from 24.19% Base to 18.84% DAPO and 17.44% VPPO, Table 9), leaving unestablished whether SPD constitutes an active failure mode on-policy.
Questions for authors
  • Constant 0.81 scaling baseline. Table 4 shows that at your default setting (m=0.05m=0.05), PACG keeps about 81% of the positive reward (advantage) on routed visual claims on average. What happens if we remove the image-masking and persistence gate entirely, and simply multiply positive advantages on all routed visual claims by a fixed constant of 0.81 (A~i,t=0.81,Ai,tbase\widetilde{A}{i,t} = 0.81 , A{i,t}^{\text{base}})? Across the same three seeds on the nine benchmarks, what accuracy do DAPO and VPPO achieve under this constant 0.81 scaling? Showing that simple uniform scaling does not match PACG's +1.8 / +1.1 gains is necessary to prove that the improvement truly comes from persistence-based gating rather than simply reducing gradient updates on visual tokens.
  • Decoupling on-policy and token likelihood breakdown: On the CrossBench4-2000 fixed responses, what are the decomposed original and masked log-likelihoods for claim tokens separated by evidence status (supported, unsupported, and explicitly contradicted), and how does corrected persistence evolve on each checkpoint's own freely generated trajectories from Table 9? Clarifying whether the rise in unsupported persistence under DAPO/VPPO is driven by contradicted claims regressing to text priors, and whether SPD persists on-policy, is necessary to validate SPD as a genuine credit assignment pathology.
3 comments
AuthorsOpenReview ↗

The response to Reviewer rTV5

Dear Reviewer,

I hope this message finds you well. We hope our responses have addressed your concerns. Your further comments are highly valuable for guiding the revision of our manuscript. Should you have any remaining questions or concerns, we welcome further discussion and look forward to your reply.

Thank you for your time and effort in reviewing our paper!

AuthorsOpenReview ↗

The response to Reviewer rTV5

Dear rTV5, With less than 10 hours left before the discussion period closes (November 18, AoE), we wanted to briefly follow up on our rebuttal. If you have a chance, we’d appreciate hearing whether our response addressed your main concerns. If anything remains unclear or unresolved, we’d be happy to discuss it further.

wish you all the best

We thank the reviewer for your time, contribution to this work and service to the community as a whole. Now that we have clarified our paper in light of the reviewer's previous questions, would you please raise your review rating or let us address more feedback?

Reviewer_xS2aOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a credit assignment problem in multimodal RL: a model can produce the correct final answer while making unsupported claims about the image, and these claims can still receive positive credit. The authors compare token probabilities for the same response under original and masked images. They find that DAPO and VPPO increase visual sensitivity but make unsupported claims more persistent under this diagnostic, whereas GRPO does not show the same deterioration.

The authors propose PACG to reduce positive credit for visual claims whose probabilities do not decrease sufficiently after masking, relative to reference tokens in the response. Experiments on Qwen-based models show improvements over DAPO and VPPO, with average accuracy gains of 1.8 and 1.1 percentage points, respectively, across nine benchmarks on the 7B model.

The authors also report improvements in their retractability metric and HallusionBench performance.

Critical strengths

The PACG can be added to both DAPO and VPPO without changing the reward or adding inference-time computation. On Qwen2.5-VL-7B, it improves average accuracy across nine benchmarks by 1.8 and 1.1 percentage points, respectively. The improvements on both parent methods provide evidence that the approach has value beyond a single training setup.

Also, the authors find the distinction between visual sensitivity and claim persistence useful. The diagnostic examines how support for individual visual claims changes under image masking, which final-answer accuracy alone does not capture. Keeping the responses fixed also allows different checkpoints to be compared on the same claims, making changes in the diagnostic results easier to interpret.

Critical weaknesses

Paper Theory: My main concern is whether PACG reduces positive credit more for unsupported claims than for supported ones. A model can misread an image and still be sensitive to masking, so an incorrect claim may retain full credit. Fig. 13 does not separate these groups, and Appendix B.2 acknowledges that the relevant assumptions remain unverified. I would like to see advantage-weighted credit retention by support status, together with a random-gate control matched for the amount of credit removed.

Experiment Setup:

Several results and experimental details also need clarification:

1)DAPO+PACG’s overall corrected persistence is −0.2558 in Fig. 4/Table 16, −0.165 in Table 14, and −0.23 at the default margin in Table 18. These values need to be reconciled, particularly since Table 14 states that it uses the same reference and checkpoint convention.

  1. Table 9 reports 2.361 direct claims per response for DAPO, while Table 10 lists 34,360 claims across 15,999 responses, giving approximately 2.148. With only one missing annotation and the stated span limit, the averaging cannot explain this gap.
Questions for authors

I would appreciate further clarification on the following questions to help me better understand and assess the paper.

  1. Could the authors report the credit retention separately for supported and unsupported claims on a sample of actual training rollouts? I would like to know whether unsupported claims receive a larger proportional reduction in positive credit.

  2. Why does DAPO+PACG’s overall corrected persistence differ between Fig. 4/Table 16 (−0.2558), Table 14 (−0.165), and Table 18 at the default margin (−0.23)? Do these values use different samples, or checkpoints?

  3. What reference-KL settings were used for each parent/PACG pair? Table 19 lists reference-log-probability computation for DAPO but not PACG. Does this reflect computation reuse, timing conventions, or a configuration difference?

1 comment
AuthorsOpenReview ↗

Response to Reviewer xS2a

Dear Reviewer,

I hope this message finds you well. We hope our responses have addressed your concerns. Your further comments are highly valuable for guiding the revision of our manuscript. Should you have any remaining questions or concerns, we welcome further discussion and look forward to your reply.

Thank you for your time and effort in reviewing our paper!

Reviewer_61x7OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies a credit-assignment problem in multimodal reinforcement learning: unsupported visual claims can receive positive reinforcement when they appear within responses that produce correct final answers. Through fixed-rollout counterfactual scoring, the authors distinguish visual sensitivity from claim persistence and show that DAPO and VPPO can improve the former while worsening the persistence of unsupported claims. They propose Persistence-Aware Credit Gating (PACG), which attenuates positive token-level advantages for visual claims that remain persistent under image corruption. Evaluation includes nine reasoning benchmarks, multiple model configurations, HallusionBench, component ablations, and a free-generation audit.

Critical strengths
  • I think the paper studies a real problem with answer-level rewards. A model can reach the correct answer while making an unsupported statement about the image along the way, and that statement can receive positive reinforcement together with the rest of the response. The reported 27.81% rate among answer-correct Base trajectories shows that this is not just a hypothetical concern.

  • I like the relatively simple training modification. PACG keeps the task reward and the parent optimizer, but reduces positive advantages on selected visual claims. It does not weaken negative advantages or require supported/unsupported labels to compute the gates. Although it adds training work, it does not require an extra scoring procedure when the trained model is used.

  • The experiments cover several useful comparisons: nine benchmarks, different scales and generations within Qwen, and ablations of claim routing, residual correction, and positive-only gating. The main 7B improvements are evaluated across three training seeds. I also appreciate the free-generation audit on uncorrupted images, since an improvement in the paper’s own persistence score would not, by itself, show that the model produces fewer unsupported claims.

Critical weaknesses
  • My main concern is whether random masking provides a reliable basis for deciding which visual claims should receive less positive credit. The corruption can affect several parts of the image at once, without checking whether the evidence supporting each claim has actually been removed. A claim’s likelihood may therefore change because of relevant evidence loss or unrelated visual changes. Residual correction reduces shared likelihood shifts, but does not distinguish these causes. I am not yet convinced that the resulting gates consistently reduce reinforcement of the claims the method intends to target.

  • I am also concerned about using fixed-response likelihood changes for credit assignment in reasoning tasks. Masking the image may weaken an early visual premise and require changes throughout the subsequent reasoning, rather than only in the tokens describing the affected object. However, PACG retains the original response prefixes when scoring each claim, so later predictions are still conditioned on premises that may no longer be supported by the image. I understand why this makes the comparison more controlled, but the paper does not sufficiently establish whether the resulting scores remain reliable for assigning credit under these conditions. In my view, this step is still somewhat heuristic, despite the reported performance gains.

  • The evidence provided in experiments seems to be stronger for average benchmark improvement than for resolving a specific, important failure of existing methods. The free-generation audit does report fewer unsupported claims, which is useful evidence. However, it remains unclear which types of errors are reduced and whether PACG systematically corrects the visual judgments that the parent models get wrong. A more focused analysis would make the practical value of the method clearer.

  • The additional training cost also needs to be considered alongside these gains. The reported 7B timing setup requires 57.3% more time per step, while the main average-accuracy improvements are 1.8 and 1.1 percentage points. I do not consider these gains negligible, but without an equal-compute comparison, it is difficult to tell whether PACG offers a better use of the training budget than further training with the parent method.

Questions for authors
  • Which experiment directly shows that “PACG reduces the fraction of answer-correct responses containing at least one unsupported direct visual claim”?

  • Was the margin of 0.05 selected using a separate validation split, or using the MMK12 test results reported in Table 4?

Reviewer_8o7hOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates a credit-assignment problem in multimodal reinforcement learning: an answer-correct reasoning trajectory can contain unsupported visual claims that are nevertheless reinforced by outcome-level rewards.

Two measurements are introduced: Evidence-Function Sensitivity (EFS), measuring how token prediction distributions change under visual intervention, and corrected claim persistence, measuring whether the likelihood of a specific visual claim decreases relative to other response tokens. Their experiments reveal that DAPO and VPPO increase visual sensitivity while making unsupported claims more persistent.

The authors propose Persistence-Aware Credit Gating (PACG), which selectively attenuates positive token advantages for persistent visual claims.

Critical strengths

The problem is well-motivated. It identifies a meaningful limitation of outcome-level RL: achieving a correct final answer does not guarantee that all intermediate reasoning statements are valid.

Separating changes in a model's predictive distribution from changes in the likelihood of a specific claim is conceptually valuable. The experimental observation that higher EFS does not necessarily imply better unsupported-claim retractability is a noteworthy finding.

Critical weaknesses

Methodological concern. The paper defines corrected persistence using the relative change in token log-likelihood between the original and degraded images (Eq 4–5). However, this measures the behavior of a fixed claim under image perturbation, not whether a model can recognize and retract an unsupported assertion. Two problems arise: First, the intervention may not remove the evidence associated with the specific claim. The random masking process may preserve relevant features while deleting unrelated information. Second, the likelihood of an unsupported claim may decrease for reasons unrelated to its factual correctness, such as general distributional shifts, image degradation, or changes in contextual confidence. Although residual correction addresses trajectory-wide likelihood drift, it cannot completely eliminate these effects. Consequently, PACG could penalize valid visual claims that are relatively insensitive to a particular intervention, while retaining unsupported claims that happen to be sensitive.

The authors argue that unsupported visual claims in correct trajectories inherit positive credit. However, in group-relative RL, a correct response does not necessarily receive a positive advantage. For example, a group in which every response is correct can have zero normalized advantage. Therefore, the reported 27.81% prevalence of unsupported claims among answer-correct trajectories does not directly measure how frequently those claims are positively reinforced during training.

Questions for authors
  1. The method does not investigate whether claim-specific negative credit could provide stronger correction than simply reducing positive credit, particularly in complex reasoning trajectories.

  2. How is sensitivity verified in experiments? It is not independently verified against visual grounding. The authors demonstrate sensitivity according to their own EFS definition, but do not sufficiently establish the relationship between that measurement and actual visual evidence utilization or claim correction.

  3. Why the entropy penalty is defined as 0.1 in Sec 5.1? I understand that Appendix A.1 clarifies that this is a positive entropy penalty, not an entropy bonus, and differs from the VPPO reference coefficient of 0.06. There is no corresponding coefficient ablation. This is particularly relevant because entropy suppression could influence model confidence and claim persistence.

  4. The backbones are quite outdated. The authors show DAPO+PACG improvements on 7B, 32B, and Qwen3-VL-2B. Such incomplete optimizer/backbone combinations weaken its generalization ability. For example, in the main Table 1 (it omits this composition and the Table 15 appendix reports only point estimates without the same multi-seed analysis), GRPO+PACG is not reported. Even further, there is no VPPO or VPPO+PACG evaluation for Qwen2.5-VL-32B in Table 1.

  5. Degradation of the image is written in 4.2 but only explained in Appendix without a clear reference. Furthermore, the authors should clarify whether degradation actually removes the evidence needed to judge a particular claim.

  6. Minors: defined q and I in Eq 1.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 8o7h

Dear Reviewer,

I hope this message finds you well. We hope our responses have addressed your concerns. Your further comments are highly valuable for guiding the revision of our manuscript. Should you have any remaining questions or concerns, we welcome further discussion and look forward to your reply.

Thank you for your time and effort in reviewing our paper!

AuthorsOpenReview ↗

Response to 7mwt

For the second concern in your weakness comment regarding the interpretation of the 27.81% statistic, we wish to provide further clarification on our rationale for reporting this value. Thank you.


Weakness2. Does Answer Correctness Imply Positive Reinforcement of Unsupported Claims?

We sincerely thank the reviewer for pointing out this critical logical distinction. We fully agree that under group-relative RL, a correct final answer does not necessarily yield a positive advantage for that trajectory. For instance, normalized advantage can be zero when every response within a group is correct. Accordingly, we will clarify this point in the revised manuscript to avoid equating answer correctness directly with positive credit assignment.

First, we would like to clarify the original intention behind the reported statistic of 27.81%. Across four multimodal reasoning benchmarks, we observe that among 8,521 trajectories with correct final answers generated by Qwen2.5-VL-7B, 2,370 trajectories (27.81%) contain at least one direct visual claim lacking supporting visual evidence.

This statistic is not intended to prove that 27.81% of these trajectories receive positive advantage during training. Instead, it reveals a pervasive phenomenon worthy of investigation: a correct final answer does not guarantee that every intermediate visual claim in the reasoning process is grounded in image evidence.

Motivated by this observation, we formulate the core research question of our work: When trajectories containing unsupported visual claims receive positive credit in RL, are these local claims also improperly reinforced? Can such training make the model more sensitive to visual inputs while making it more likely to persist in evidence-free visual claims?

This is the starting point of our investigation into Sensitivity–Persistence Decoupling (SPD). Intuitively, stronger visual sensitivity should make the model more prone to retract incorrect claims when visual evidence is perturbed. Counterintuitively, our experiments show that while DAPO and VPPO improve EFS, the corrected persistence of unsupported claims can degrade.

Second, to verify whether this motivation aligns with real positive credit exposure during training, we add an Online Positive-Credit Exposure Audit. We analyze the real training rollouts of Qwen2.5-VL-7B under GRPO, DAPO and VPPO (steps 181–200, seed 0), and quantify how often unsupported direct visual claims receive positive advantage in optimization.

OptimizerUnsupported SpansPositive-Advantage ExposureNormalized Positive-Credit Mass
GRPO1,84248.6%0.29
DAPO1,59667.4%0.81
VPPO2,10578.1%0.94

The results demonstrate that unsupported visual claims can indeed receive substantial positive credit in practice. Specifically, DAPO and VPPO achieve positive-advantage exposure rates of 67.4% and 78.1%, respectively, showing that group-relative advantage does not automatically prevent unsupported visual claims from being positively reinforced.

Furthermore, the positive-credit mass increases across GRPO, DAPO, and VPPO, following the same ordering as the degradation of unsupported persistence. That said, given other confounding differences between optimizers, we treat this as correlational evidence supporting our motivation, rather than standalone causal proof.

In summary, we highlight three separate layers of conclusions:

  1. The 27.81% figure shows that trajectories with correct answers but unsupported visual claims are common.
  2. The Online Credit Exposure Audit directly confirms that such unsupported claims can receive positive credit during RL training.
  3. The SPD analysis further studies the relationship between positive credit assignment, visual sensitivity, and claim persistence, identifying notable behavioral shifts under DAPO/VPPO.

Therefore, our core claim is not that all correct answers reinforce incorrect visual claims. Instead, we argue that when unsupported visual claims receive positive credit, existing RL optimization may fail to sufficiently distinguish final answer correctness from the evidential reliability of local visual claims. This credit-assignment failure merits dedicated study.

In the revised manuscript, we have explicitly separated claim prevalence, positive-credit exposure, and post-training persistence, and state their logical relationships more precisely. This improves the readability and rigor of our introduction. Thank you!

We thank the reviewer again for this valuable comment. It helps us formulate our research motivation and mechanistic assumptions with greater rigor.

We wish you all the best.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

For clarity, we refer to Reviewers 8o7h, 61x7, xS2a, rTV5, and WnSx as R1–R5, respectively.

We sincerely thank all reviewers for their thoughtful and constructive feedback. We have incorporated additional experiments, methodological clarifications, and reporting corrections into the revised manuscript. We have also released an anonymous reference implementation at PACG Code, with its scope clarified in the Reproducibility Statement.


Core Contributions of Our Work

  1. Sensitivity–Persistence Decoupling (SPD): Our fixed-rollout diagnostic separates visual sensitivity from claim persistence. We find that DAPO and VPPO can increase visual sensitivity while making unsupported claims more persistent under intervention; GRPO does not exhibit the same deterioration.

  2. Persistence-Aware Credit Gating (PACG): We combine direct-claim routing, residual-calibrated persistence, and positive-only gating to selectively attenuate credit for persistent visual claims, without supported/unsupported labels for gating or additional inference computation.

  3. Task-Level and Behavioral Validation: PACG improves nine-benchmark average accuracy by 1.8/1.1 percentage points over DAPO/VPPO on Qwen2.5-VL-7B. Additional model evaluations and behavioral audits support its effectiveness and statistical selectivity.


Updates to Experimental Results during Rebuttal

  • Main Text and Appendix E.7: Added evidence-disruption audits, claim-targeted occlusion, and free-continuation validation linking corrected persistence to semantic retention (R1, R2, R5). Added on-policy retention experiments using each model’s own unsupported claims (R4).

  • Main Text and Appendix E.4: Added actual training-rollout audits of positive-credit exposure and advantage-weighted retention. Supported–unsupported retention gaps reach 13.8/27.6 percentage points under DAPO/VPPO+PACG (R1–R3).

  • Main Text and Appendix F.5: Added three-seed constant-scaling and credit-mass-matched random-gating controls to test PACG beyond generic credit attenuation (R3, R4).

  • Main Text and Appendix D.6: Added answer-correct conditional statistics: unsupported-response incidence decreases from 21.8% to 18.6% for DAPO and 22.4% to 18.9% for VPPO (R2).

  • Main Text and Appendix F.1: Added 32B VPPO/PACG comparisons and three-seed GRPO/PACG evaluations on 7B and 32B (R1).

  • Appendices F.4–F.7: Extended margin sensitivity analysis (R2), added active negative-credit controls (R1), and included an unofficial matched-setting SRPO reproduction with explicit limitations (R5).

  • Main Text and Appendix H.1: Added an equal-wall-clock comparison. At approximately 40.6 hours, PACG achieves 59.9% average accuracy versus 59.2% for extended DAPO, together with lower unsupported persistence (R2).


Updates to In-Depth Discussions during Rebuttal

  • Abstract, Introduction, and Preliminaries: Clarified the progression from unsupported claims to the credit-assignment question and SPD diagnosis. Distinguished the 27.81% prevalence statistic from actual positive-credit exposure (R1).

  • Main Text and Appendices E.4/E.7: Clarified the distinctions among sensitivity, persistence, factual support, and semantic retraction; explained intervention and fixed-prefix limitations; and distinguished statistical selectivity from per-claim guarantees (R1–R3, R5).

  • Main Text and Appendix J: Distinguished unsupported-claim generation frequency from conditional persistence and refined the interpretation of PACG’s performance gains (R4).

  • Related Work and Appendix F.7: Expanded the existing SRPO discussion to acknowledge shared technical elements and clarify PACG’s distinct contributions (R5).

  • Appendices A.2, F.4, and H: Clarified regularization settings and margin selection, corrected training overhead to 64.5%, and reconciled affected persistence and claim-count statistics (R1–R3).

All additions and revisions are highlighted in blue for ease of reference.Our last manuscript is updated on Nov 14, AOE time.

We look forward to the reviewers' favourable consideration and remain grateful for their valuable feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.