Shared Semantics, Divergent Mechanisms: Unsupervised Feature Discovery by Aligning Semantics and Mechanisms

As large language models are increasingly deployed in high-stakes settings, there is a growing need for tools that audit not only model outputs but also the internal computations that produce them. Circuit analysis is a central approach in mechanistic interpretability, but it is typically target-conditioned, explaining a single prompt paired with a chosen completion. This target-conditioned setup can obscure heterogeneity across a model's continuation distribution. We introduce distribution-level unsupervised feature discovery, which clusters sampled continuations using both semantic content and sequence-level mechanistic attributions, without manually specifying target outputs. Our method represents each continuation with a semantic embedding and a prefix-to-continuation attribution signature, then optimizes a rate-distortion objective that trades off semantic coherence, mechanistic consistency, and cluster granularity. Across clustering and steering analyses, the discovered clusters expose continuation modes that single-view baselines miss and provide interventional evidence that cluster signatures correspond to actionable mechanistic factors. Overall, our approach complements circuit analysis and behavioral evaluation by providing a scalable audit of the mechanisms underlying a model’s continuation distribution.

AI language models are widely deployed where their decisions impact real consequences, so we need to check not just what they say but how they reach their answers. The usual way of inspecting a model's inner workings looks at a single response to a single query—but a model can answer in many different directions, and studying just one hides that variety. We developed a method that examines the full range of a model's possible responses. It divides them into groups based on two things at once: what each response means, and the internal "fingerprint" of how the model produced it. The result reveals distinct response patterns that simpler approaches miss. And when we adjust the model's inner workings, its behavior changes as our groupings predict, which can be viewed as evidence that these patterns are not mere correlations, but a handle for controlling the direction of responses. In short, our approach offers a scalable way to audit the machinery behind everything a model might say.