Evidence-RL: Towards Evidence-intensive Visual Reasoning

2026-08-08

AI Summary

Current vision-language models often rely on linguistic priors or irrelevant visual cues rather than genuine evidence from images for reasoning. This work proposes Counterfactual Evidence Disentanglement (CED), a novel approach that introduces, for the first time, a causally informed evidence auditing mechanism during training. CED leverages weakly supervised object proposals to identify evidence regions and combines a counterfactual region neutralization strategy with a GRPO reinforcement learning framework to construct an evidence-sensitive reward signal—without requiring question-level evidence annotations. The method incurs no additional inference overhead while effectively guiding the model toward authentic, object-centric evidence pathways. Extensive experiments demonstrate that CED consistently outperforms existing reinforcement learning–based post-training methods across nine public benchmarks and four backbone architectures, confirming the efficacy of its evidence-oriented modeling paradigm.

Abstract

Vision-Language Models (VLMs) should answer from concrete image evidence rather than language priors, dataset shortcuts, or irrelevant visual context. Existing perception-aware post-training methods encourage image use through global perturbations or attention proxies, but they do not test whether a sampled answer causally depends on the local evidence that supports it. We propose Counterfactual Evidence Disentanglement (CED), a training-time evidence audit for VLM grounding. For each response, CED neutralizes an object-centric Evidence Region and compares the resulting support drop against matched non-evidence Regions. We combine this signal with answer correctness inside GRPO, rewarding correct answers that rely on the evidence path rather than shortcut or nuisance paths. CED uses weak object-level proposals, requires no question-specific evidence annotations, and adds no inference-time overhead. Across nine public benchmarks and four backbones, CED outperforms prior RL-based post-training methods, with targeted analyses verifying its object-centric signal.

Problem

• visual reasoning
• evidence grounding
• vision-language models
• language priors
• dataset shortcuts

Innovation

• Counterfactual Evidence Disentanglement
• Vision-Language Models
• Evidence Grounding
• Reinforcement Learning
• Object-centric Reasoning