Thinking with Deltas: Incentivizing Reinforcement Learning via Differential Visual Reasoning Policy

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced reasoning capabilities in Large Language Models. However, adapting RLVR to multimodal domains suffers from a critical $\mathit{\text{perception-reasoning}}\mathit{\text{decoupling}}$. Existing paradigms, driven by text-centric outcome rewards, reasoning in language medium, inadvertently encourage models to bypass visual perception. We empirically validate this through blind experiments: state-of-the-art policies maintain or surprisingly improve performance even when visual inputs are entirely removed. This reveals that these models degenerate into $\mathit{\text{blind}}\mathit{\text{reasoners}}$, exploiting linguistic priors to generate plausible answers instead of attending to visual evidence. In response, we propose $\mathbf{\text{Thinking}}\mathbf{\text{with}}\mathbf{\text{Deltas}}$, a framework driven by a $\mathbf{\text{Differential}}\mathbf{\text{Visual}}\mathbf{\text{Reasoning}}\mathbf{\text{Policy}}\mathbf{\text{(DVRP)}}$. DVRP introduces intrinsic supervision via visual triplets, comprising original, masked, and perturbed inputs. It optimizes the model to maximize reasoning divergence from masked inputs (enforcing $\mathit{\text{visual}}\mathit{\text{sensitivity}}$) while minimizing divergence from perturbed inputs (ensuring $\mathit{\text{visual}}\mathit{\text{robustness}}$). By aligning reasoning variations strictly with the $\mathit{\text{Delta}}$ of visual information, DVRP inherently bolsters visual understanding capabilities and significantly outperforms state-of-the-art methods on both general and medical benchmarks, without requiring external annotations or auxiliary tools.