DeepLook: Uncertainty-Guided Branching with Lookahead Confidence for Efficient Reasoning
DeepLook is a training-free method that improves large language model reasoning by focusing computational effort only where uncertainty is high. It monitors segment confidence during decoding and branches only at these uncertainty bottlenecks, ranking potential continuations using Average Lookahead Confidence (ALC) before selecting the best path. This approach achieves better accuracy with significantly fewer tokens compared to standard scaling methods.