Not All Tokens See Equally: Perception-Grounded Policy Optimization for Large Vision-Language Models

Not All Tokens See Equally: Perception-Grounded Policy Optimization for Large Vision-Language Models explores A fine-grained credit assignment framework for vision-language models that amplifies learning signals for visually-dependent tokens, improving multimodal reasoning performance.. Commercial viability score: 7/10 in LLM Fine-tuning.