PEAR is an adaptive advantage reshaping method designed to improve perception and exploration in large vision-language models. Effective multimodal reasoning requires a model to both reliably exploit visual information and sufficiently explore diverse reasoning paths. Previous work combined visual dependency and token entropy to reshape sequence-level advantage, but scaling entropy by visual dependency can underweight high-entropy tokens with low visual dependency, even though they are important for exploration. PEAR addresses this limitation by keeping visual dependency as the base signal, adding an entropy-based exploration gain only at high-entropy positions, with lower visual dependency yielding a larger gain. The fused signal is then mapped to positive token-level weights that reshape the sequence-level advantage. Experiments on Qwen2.5-VL-3B/7B show that PEAR achieves the highest average performance among baselines on multimodal reasoning benchmarks and also leads on textual reasoning and general multimodal benchmarks.
Review