DeepLook Improves LLM Reasoning Efficiency with Selective Lookahead.

Improving Large Language Model (LLM) reasoning often involves inference-time scaling, which can yield substantial gains on difficult tasks. However, current approaches are often inefficient in how they allocate computational resources during a reasoning trace. Many reasoning failures show early signs of uncertainty before a definitive wrong answer is produced. DeepLook, a new training-free monitor-and-intervene decoding framework, addresses this inefficiency by strategically focusing lookahead compute where uncertainty is highest. It aggregates token-level confidence into segment-level signals, triggering a fixed-horizon lookahead exploration when confidence drops. Candidate continuations are then ranked using Average Lookahead Confidence (ALC), which averages segment-level confidence over rollouts, before pruning and aggregating through voting. This selective, future-aware intervention significantly improves the accuracy-token cost trade-offs compared to uniformly scaling complete reasoning trajectories. Experiments on mathematics benchmarks across various LLMs demonstrated that DeepLook boosts accuracy while drastically reducing token generation, showcasing its effectiveness in making LLM reasoning more efficient and powerful.