How restarting reasoning at different points improves large language model answers
Expanding LLM Reasoning
Machine LearningArtificial IntelligenceComputation and Language
Summary
When large language models (LLMs) work through a problem, they often create chains of reasoning steps. This paper finds that instead of always starting these chains from the beginning, restarting them at specific intermediate steps can make the final answer more accurate. The authors measure how useful each restart point is and identify strategies that beat simply choosing restarts evenly or always restarting from the end. They also show that choosing the best restart points is still an open challenge for future work. Lastly, they reveal that small technical details, like how to break ties in scores, can significantly affect these improvements.
What this means in practice
- •For large language model engineers: Design model inference workflows that improve accuracy by selectively restarting reasoning chains at effective intermediate steps.
- •For ai product teams: Enhance AI-assisted problem-solving tools by applying smarter restart strategies that increase answer correctness without extra output generation.
Authors
Rian Atri, Evan Luo
Abstract
Extra inference compute is usually spent on sampling more reasoning chains. We study where inside an existing chain an additional continuation should begin. We define expansion utility, the change in correctness from restarting a chain at a stored step, and measure it at every eligible step for nine models on six benchmarks (41 model and benchmark cells). Restart position matters: steps selected on one set of continuations beat uniform placement when scored on disjoint ones, in held-out audits on 5, 16, and 38 cells (+4.25 points [+2.51, +6.63] in a fresh five-cell audit). A fixed rule that restarts from the last eligible steps, always-last, is a strong baseline: our learned router beats uniform placement but shows no detected gain over it, and on DeepSeek-R1-Distill-Qwen-14B/MATH-500 always-last exceeds the exact self-consistency frontier at matched aggregate generated output by +0.052 [+0.008, +0.098], using 0.774x the aggregate generated output of four-sample self-consistency. Cross-fitted oracle selection still finds held-out headroom beyond declared positional classes, a target for future selectors. Finally, breaking step-label ties by earliest index flips the sign of a pointwise selector's gain over uniform placement in every seed of a five-seed diagnostic with four rollouts per step; randomized ties remove the bias.