DeepLook introduces a monitor-and-intervene decoding framework that focuses computational resources on uncertainty bottlenecks during inference. It aggregates token-level confidence into segment-level signals, triggering when confidence drops relative to recent history, and explores candidate continuations with fixed-horizon lookahead. Branches are ranked by Average Lookahead Confidence (ALC), which is the average segment-level confidence over rollout continuations, followed by pruning and voting aggregation.
On four competition-style mathematics benchmarks, including DeepSeek-R1-8B, Qwen3-32B, GPT-OSS-20B, and GPT-OSS-120B, DeepLook shifts the accuracy-token-cost Pareto frontier. It improves accuracy in 11 out of 16 settings while reducing dataset-level token generation by an average of 87.3%, with gains of +3.1 on AIME25 with Qwen3-32B and +8.8 on BRUMO25 with GPT-OSS-20B. These results indicate that selective, future-aware interventions yield significantly stronger accuracy-cost trade-offs compared to uniformly scaling complete reasoning trajectories.
Code is available here.