Entropy-Gradient Inversion: Moving Toward Internal Mechanism of Large Reasoning Models

In three linesResearchers identify Entropy-Gradient Inversion, a negative correlation between token entropy and logit gradients, as a geometric fingerprint of Large Reasoning Models' reasoning capability. They propose Correlation-Regularized Group Policy Optimization (CorR-PO), embedding this inversion signature into RL reward regularization, outperforming baselines across multiple reasoning benchmarks.

Summary generated by Claude — human-verified