Does Seeing More Mean Knowing More? Mono-Anchored Advantage Normalization for Multi-Source Visual Reasoning

Researchers introduced MARS, a mono-anchored framework for multi-source visual reasoning.

It treats each modality as an independent source and uses mono-source rewards as dynamic anchors in advantage normalization.

This helps estimate information gain, reduce cross-modal interference, and decide when extra inputs like infrared or depth actually help.

The approach improved reinforcement learning with verifiable rewards on GRPO and DAPO benchmarks.