Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Fangan Dong, Zuming Yan, Xuri Ge, Zhiwei Xu, Mengqi Zhang, Xuanang Chen, Ben He, Xin Xin, Zhumin Chen, Ying Zhou

TL;DR

This work addresses the bottleneck of unreliable reasoning in LLMs by introducing AdaRAS, a test-time activation-steering framework that identifies a sparse set of Reasoning-Critical Neurons (RCNs) and adaptively alters their activations to correct faulty reasoning trajectories. RCNs are discovered via a polarity-aware mean-difference criterion between correct and incorrect reasoning traces, and steering is gated by an adaptive failure predictor to avoid degrading already-correct cases. AdaRAS demonstrates consistent gains across ten math and coding benchmarks, outperforming some post-training approaches and generalizing across datasets and larger models, with minimal computation cost. Mechanistic analyses show AdaRAS stabilizes latent reasoning trajectories while preserving semantic representations, supporting a practical, plug-and-play path to more reliable LLM reasoning.

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.

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

TL;DR

This work addresses the bottleneck of unreliable reasoning in LLMs by introducing AdaRAS, a test-time activation-steering framework that identifies a sparse set of Reasoning-Critical Neurons (RCNs) and adaptively alters their activations to correct faulty reasoning trajectories. RCNs are discovered via a polarity-aware mean-difference criterion between correct and incorrect reasoning traces, and steering is gated by an adaptive failure predictor to avoid degrading already-correct cases. AdaRAS demonstrates consistent gains across ten math and coding benchmarks, outperforming some post-training approaches and generalizing across datasets and larger models, with minimal computation cost. Mechanistic analyses show AdaRAS stabilizes latent reasoning trajectories while preserving semantic representations, supporting a practical, plug-and-play path to more reliable LLM reasoning.

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.
Paper Structure (59 sections, 8 equations, 7 figures, 10 tables)

This paper contains 59 sections, 8 equations, 7 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Task Formulation
  4. LLM Reasoning Signatures
  5. Finding 1. LLM reasoning traces leading to correct versus incorrect answers exhibit distinct activation patterns.
  6. Finding 2. Token-level neuron activations are predictive of the final correctness of LLM reasoning.
  7. Methodology
  8. Reasoning Neuron Identification
  9. Critical Activation Selection
  10. Adaptive Intervention Strategy
  11. Contrastive Data Construction
  12. Experiments
  13. Experimental Setup
  14. Dataset.
  15. Baseline.
  16. ...and 44 more sections

Figures (7)

  • Figure 1: An example of activation steering correcting an erroneous reasoning trajectory.
  • Figure 2: Comparison of activations of key neurons under successful and failed reasoning on AIME.
  • Figure 3: Overview of AdaRAS: (1) reasoning neuron identification (\ref{['sec:method:identify']}), which identifies critical neurons by measuring global activation differences between contrastive reasoning trajectories; (2) critical activation selection (\ref{['sec:method:steer']}), which further refines RCNs based on activation polarity variations; (3) adaptive intervention (\ref{['sec:method:adaptive']}), which enhances the reliability of steering by predicting reasoning failures and performing adaptive interventions.
  • Figure 4: Effect of hyperparameter.
  • Figure 5: Visualization of activation shifts induced by steering. Neurons are sorted by their indices in layer, and unsteered neurons are omitted for clarity.
  • ...and 2 more figures