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

Benchmark Model Rank Results
visual-question-answering-vqa-on-simplevqaClaimDiff-RL (actor-only) (init: Qwen3VL-32B-Instruct)#8Accuracy: 53.2
visual-question-answering-vqa-on-blinkClaimDiff-RL (actor-only) (init: Qwen3VL-32B-Instruct)#16Accuracy: 66.1
visual-question-answering-vqa-on-realworldqaClaimDiff-RL (actor-only) (init: Qwen3VL-32B-Instruct)#46Accuracy: 71.1