Entropy Centroids as Intrinsic Rewards for Test-Time Scaling

A new method uses entropy centroids as intrinsic rewards to scale test-time compute for large language models, avoiding external reward models. The approach leverages the observation that high-entropy tokens cluster into consecutive groups during inference, providing stable uncertainty signals. This temporal structure is formalized into segment-level rewards, offering an alternative to confidence-based or entropy-based methods that suffer from noise. The work is published on arXiv under ID 2604.26173.