k-Sparse Autoencoders Reveal Reasoning Model Features

A LessWrong community post presents an early-stage project investigating whether k-sparse autoencoders (k-SAEs) can extract interpretable features from a small reasoning model. K-sparse autoencoders enforce sparsity by retaining only the k largest magnitude activations in the latent layer, producing sparser and more targeted feature decompositions than standard sparse autoencoders trained with an L1 penalty. The project focuses on finding evidence that internal activations correspond to recognizable "thinking patterns" or discrete reasoning steps.

Sparse autoencoders have become a central tool in mechanistic interpretability. OpenAI's scaling work (Gao et al., 2024) demonstrated that SAEs trained on GPT-4 residual stream activations recover monosemantic, interpretable features at scale. A peer-reviewed paper (arXiv 2503.18878) specifically applied sparse autoencoders to reasoning traces in large language models, reporting features that activate during uncertainty, reflection, and exploratory thinking. The LessWrong project aims to test similar ideas using the k-sparse variant on a smaller reasoning model, which lowers the compute barrier and may clarify whether such features generalize to smaller-scale reasoning architectures.

The post is an early community project, not a completed or peer-reviewed study. The snippet indicates the work is framed as an investigation rather than a findings report, so claims about what the autoencoders actually reveal are not yet confirmed. The single-source LessWrong format means verification against primary experimental data is not currently possible.

If k-SAEs can reliably surface reasoning features in small models, the approach could become a lightweight diagnostic tool - practitioners could apply it to understand chain-of-thought structure, identify reasoning failure modes, or steer model behavior without requiring large-scale compute. The result, positive or negative, informs the broader question of whether sparse dictionary learning generalizes across model scale and architecture in reasoning-focused settings.