Identifying and Transferring Reasoning-Critical Neurons: Improving LLM Inference Reliability via Activation Steering

TL;DR

This paper introduces AdaRAS, a lightweight framework that improves large language model reasoning reliability by selectively steering activations of critical neurons during inference, leading to significant performance gains without extra training.

Contribution

The paper identifies reasoning-critical neurons in LLMs and proposes AdaRAS, a novel activation steering method that enhances reasoning accuracy during inference without additional training.

Findings

Achieves over 13% improvement on AIME-24 and AIME-25 benchmarks.

Demonstrates transferability across datasets and scalability to stronger models.

Outperforms post-training methods without extra training or sampling costs.

Abstract

Despite the strong reasoning capabilities of recent large language models (LLMs), achieving reliable performance on challenging tasks often requires post-training or computationally expensive sampling strategies, limiting their practical efficiency. In this work, we first show that a small subset of neurons in LLMs exhibits strong predictive correlations with reasoning correctness. Based on this observation, we propose AdaRAS (Adaptive Reasoning Activation Steering), a lightweight test-time framework that improves reasoning reliability by selectively intervening on neuron activations. AdaRAS identifies Reasoning-Critical Neurons (RCNs) via a polarity-aware mean-difference criterion and adaptively steers their activations during inference, enhancing incorrect reasoning traces while avoiding degradation on already-correct cases. Experiments on 10 mathematics and coding benchmarks demonstrate consistent improvements, including over 13% gains on AIME-24 and AIME-25. Moreover, AdaRAS exhibits strong transferability across datasets and scalability to stronger models, outperforming post-training methods without additional training or sampling cost.