#natural language process...
Jan 2026
Identifying and Transferring Reasoning-Critical Neurons: Improving LLM Inference Reliability via Activation Steering
AdaRAS (Adaptive Reasoning Activation Steering), a lightweight test-time framework that improves reasoning reliability by selectively intervening on neuron activations during inference, enhancing incorrect reasoning traces while avoiding degradation on already-correct cases.
Fang Dong, Zu-Ming Yan, Xuri Ge et al.
· arXiv.org · 5 citations