ClaimDiff-RL: Fine-Grained Caption Reinforcement Learning through Visual Claim Comparison

Researchers introduced ClaimDiff-RL, a fine-grained reinforcement learning method for image captioning.

It uses reference-conditioned atomic claim differences as reward signals, separately tracking hallucinated and omitted claims.

The approach improves the balance between faithfulness and coverage, reducing factual errors without losing detail.

ClaimDiff-RL performs strongly across diagnostic, captioning, and VQA benchmarks, including multimodal-judge evaluations.