Prioritizing the Best: Incentivizing Reliable Multimodal Reasoning by Rewarding Beyond Answer Correctness

Develop an RL library for multimodal reasoning that uses 'trajectory supervision' instead of just final answer reward. This helps developers build models that actually reason through visual tasks.

Suggested repo: vision-chain-rl

"Reward the thought process, not just the final guess."

Estimated effort: 120h