Faithful-MR1: Faithful Multimodal Reasoning via Anchoring and Reinforcing Visual Attention

In three linesFaithful-MR1 is a training framework for MLLMs improving multimodal reasoning via reinforcement learning. It anchors visual attention directly on image regions (not via textual descriptions) and reinforces faithful use through counterfactual image intervention. Results on Qwen2.5-VL-Instruct 3B/7B with substantially less training data.

Summary generated by Claude — human-verified