Visual sensitivity and claim retractability improve multimodal learning accuracy
Visual sensitivity is not claim retractability: persistence-aware credit assignment for multimodal reinforcement learning
Artificial Intelligence
Summary
Some AI models answer questions about images correctly but still make claims that the image does not actually support. The authors found that rewarding answers without checking claim accuracy lets wrong visual claims keep getting credit. They created a way to detect whether models keep supporting claims even when shown changed images, then proposed a method to reduce credit for wrong claims that don’t get retracted. This approach improved accuracy on multiple tests and made unsupported image claims less persistent.
What this means in practice
- •For multimodal ai developers: Improve training of vision-language models by reducing credit for unsupported visual claims to boost answer accuracy.
- •For automated visual reasoning teams: Use persistence-aware credit gating to diagnose and mitigate persistent unsupported claims in image-based reasoning systems.
Authors
Zhongan Bi, Kepeng Lin, Xuanang Gao, Yuhan Sun, Lianrun Zhang
Abstract
Reinforcement Learning with Verifiable Rewards (RLVR) has been extended to Large Vision-Language Models (LVLMs), and perception-aware methods further encourage policies to rely on visual evidence. Yet relying on the image does not guarantee that visual claims are supported by it. Before RL training, 27.81% of the correctly answered responses of Qwen2.5-VL-7B on four multimodal reasoning benchmarks contain at least one direct visual claim that the image does not support. Since outcome-level RL rewards each response as a whole, these claims inherit the positive credit of the correct answer. We introduce a fixed-rollout counterfactual diagnostic that re-scores the same response under an intervened image to separate Evidence-Function Sensitivity (EFS), how strongly the model's predictions change, from claim persistence, whether the model keeps supporting the same claim rather than retracting it. The diagnostic reveals Sensitivity-Persistence Decoupling (SPD): under DAPO and VPPO, EFS increases and claims become more retractable overall, yet unsupported claims become significantly more persistent, whereas GRPO raises EFS without this deterioration. We therefore propose Persistence-Aware Credit Gating (PACG), which attenuates positive credit for unusually persistent visual claims and leaves all other credit unchanged. It requires no supported/unsupported labels and adds no inference cost. On Qwen2.5-VL-7B, PACG raises the nine-benchmark average over three seeds from 58.1% to 59.9% with DAPO and from 59.8% to 60.9% with VPPO, while making unsupported claims more retractable. The gains extend to a larger model, a newer backbone, and the accuracy of HallusionBench also improves consistently. These results suggest that visual sensitivity and claim retractability are complementary dimensions of multimodal credit assignment.