Sparse Autoencoder-Based Steering for Controllable LLM Reasoning

Researchers propose SAE-Steering, a method to control reasoning strategies in Large Reasoning Models (LRMs) by leveraging Sparse Autoencoders (SAEs) to decompose entangled hidden states into a disentangled feature space. The approach addresses the inefficiency and errors of autonomous reasoning strategy selection. A two-stage pipeline identifies strategy-specific features from the vast pool of SAE features, enabling fine-grained control over reasoning processes such as backtracking and cross-verification. The work is detailed in arXiv:2601.03595.