VCap: Hypergeometric Rewards Improve Visual Captioning

A new reinforcement learning method for visual captioning, VCap, uses a Witness-Adjudicator reward to improve factual accuracy. The approach pairs a reference caption (witness) with visual signals (adjudicator) to verify factual consistency, achieving hypergeometric-distribution-level precision. This enables effective learning even from imperfect references, addressing limitations of existing reward designs that lack fine-grained factual verification. The method targets omission and hallucination in multimodal large language models (MLLMs).