Can We Break LLMs Out of Self-Loops? Fine-Grained Reasoning Control with Activation Steering
The paper says failed reasoning traces can stall in latent-state self-loops, and tests activation steering as an inference-time way out.
The authors introduce SOPHIA, which treats a model’s reasoning trace as a sequence of latent states instead of plain text. It classifies prefixes into states, tracks step-level transitions, and builds steering vectors for state-pair moves. At inference time, a controller can detect looping behavior and apply the matching vector toward a target state. The paper reports better intervention on self-loop failures, with gains in end-task accuracy and token efficiency. ArXiv · AI/CL/LG's note
The authors introduce SOPHIA, which treats a model’s reasoning trace as a sequence of latent states instead of plain text. It classifies prefixes into states, tracks step-level transitions, and builds steering vectors for state-pair moves. At inference time, a controller can detect looping behavior and apply the matching vector toward a target state. The paper reports better intervention on self-loop failures, with gains in end-task accuracy and token efficiency. ArXiv · AI/CL/LG's note
score 5