Grounding Multimodal Reasoning with Evidence-Ablated Negatives
Evidence-ablated negatives stop RLVR models from answering without reading the image. Masked patches expose shortcuts the policy then suppresses.
The claim
Evidence-ablated negatives stop RLVR models from answering without reading the image.
The problem
RLVR improves multimodal reasoning with verifiable rewards.
Outcome rewards can come from language priors instead of the image.
Good negatives must stay plausible while exposing the shortcut.
How it works
EVAN ranks examples by image information gain.
It masks the most discriminative patches of the image.
The current policy then answers from language priors alone.
Those answers enter optimization as bounded auxiliary signals.
A mixed-ranking strategy keeps the training data diverse.
Figures
What changes
Shortcut learning is penalized while RL training stays stable.
Results beat strong RLVR baselines on multimodal reasoning benchmarks.
Limits
The abstract reports no single numeric gain.
Gains are shown across benchmarks the authors selected.
The method is evaluated on benchmarks, not deployed systems.
Source · arXiv