DeepLook: Training-Free Inference Scaling for LLM Reasoning

A recent publication on arXiv presents DeepLook, a decoding framework that enhances reasoning in large language models without the need for training. This innovative method focuses computational resources on areas of uncertainty, aggregating confidence at the token level into signals at the segment level. It activates when confidence wanes and evaluates potential continuations using a fixed-horizon lookahead. Candidate branches are prioritized based on Average Lookahead Confidence (ALC) and are refined through a voting process. This strategy effectively tackles the shortcomings of current inference-time scaling techniques.