Depth-Entropy Guided Sampling for Training-Free LLM Reasoning

Short summary

DEGS (Depth-Entropy Guided Sampling) is a training-free, test-time method that exploits layer-wise entropy collapse in transformers as a quality signal for LLM reasoning. Stronger reasoners exhibit 'late collapse' where entropy stays elevated until deeper layers. By combining sequence likelihood with a per-sequence collapse depth signal inside an MCMC framework, DEGS achieves state-of-the-art training-free accuracy across three models and four benchmarks, even surpassing GRPO out-of-domain on GPQA for all three models, at single-digit-percent overhead.

• •DEGS uses layer-wise entropy collapse depth as an intrinsic quality signal for training-free test-time reasoning improvement
• •The method surpasses GRPO out-of-domain on GPQA across three models without any training or labeled data
• •Gains are largest on harder splits and out-of-domain tasks where likelihood-only sampling falls short

Generated with AI, which can make mistakes.