Reproduce “Identifying and Transferring Reasoning-Critical Neurons: Improving LLM Inference Reliability via Activation Steering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fangan Dong, Zuming Yan, Xuri Ge, Zhiwei Xu, Mengqi Zhang, Xuanang Chen, Ben He, Xin Xin and 2 more
AdaRAS steers selected MLP neurons during reasoning, applying the intervention only when a predictor expects failure. It identifies neurons whose average activations differ in sign between correct and incorrect reasoning traces, then shifts their activations at inference. On AIME-25, Qwen3-1.7B with AdaRAS reaches 54.55% accuracy versus 40.91% for CoT; results are limited to tested Qwen3 models and STEM benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.