Stabilizing Efficient Reasoning with Step-Level Advantage Selection

TL;DR

This paper introduces Step-level Advantage Selection (SAS), a method that stabilizes and improves efficient reasoning in large language models by selectively pruning low-confidence steps, leading to better accuracy and shorter reasoning traces.

Contribution

The paper presents SAS, a novel step-level advantage method that enhances reasoning efficiency and stability without sacrificing accuracy in large language models.

Findings

SAS improves Pass@1 accuracy by 0.86 points over length-aware baselines.

SAS reduces reasoning length by 16.3% on average.

Short-context post-training induces reasoning compression but can cause instability.

Abstract

Large language models (LLMs) achieve strong reasoning performance by allocating substantial computation at inference time, often generating long and verbose reasoning traces. While recent work on efficient reasoning reduces this overhead through length-based rewards or pruning, many approaches are post-trained under a much shorter context window than base-model training, a factor whose effect has not been systematically isolated. We first show that short-context post-training alone, using standard GRPO without any length-aware objective, already induces substantial reasoning compression-but at the cost of increasingly unstable training dynamics and accuracy degradation. To address this, we propose Step-level Advantage Selection (SAS), which operates at the reasoning-step level and assigns a zero advantage to low-confidence steps in correct rollouts and to high-confidence steps in verifier-failed rollouts, where failures often arise from truncation or verifier issues rather than incorrect reasoning. Across diverse mathematical and general reasoning benchmarks, SAS improves average Pass@1 accuracy by 0.86 points over the strongest length-aware baseline while reducing average reasoning length by 16.3%, yielding a better accuracy-efficiency trade-off.