New entropy-based decoding method helps LLMs reason better by focusing computation on uncertain decisions rather than confident ones

Proposes entropy-guided decoding framework that identifies high-uncertainty token positions and selectively branches on vulnerable points during generation

Addresses limitations of existing methods: greedy decoding and beam search suffer from error propagation, while sampling introduces unwanted randomness

Maintains dynamic pool of partial rollouts that expands until solutions complete, concentrating computational resources where uncertainty is greatest

Improves upon self-consistency approach which aggregates multiple rollouts but requires significant computational overhead

Applies rollout-level entropy for efficient termination, enabling practical deployment without excessive processing