Do Sparse Autoencoders Identify Reasoning Features in Language Models?

TL;DR

This work questions whether sparse autoencoders isolate genuine reasoning features in large language models, proposing a falsification framework that combines causal token injection and LLM-guided counterexamples. Across 20 configurations, token-level cues account for the majority of features identified by contrastive methods, and remaining context-dependent features are systematically falsified as confounds rather than core reasoning; steering yields only minor performance changes. By formalizing a rigorous criterion set for genuine reasoning features and demonstrating the dominance of lexical correlates, the study highlights the risk of over-interpreting SAE directions without causal validation. The findings advocate for falsification-based evaluation in mechanistic interpretability and underscore the need for richer representations to capture complex, distributed reasoning if it exists in LLMs.

Abstract

We investigate whether sparse autoencoders (SAEs) identify genuine reasoning features in large language models (LLMs). Starting from features selected using standard contrastive activation methods, we introduce a falsification-oriented framework that combines causal token injection experiments and LLM-guided falsification to test whether feature activation reflects reasoning processes or superficial linguistic correlates. Across 20 configurations spanning multiple model families, layers, and reasoning datasets, we find that identified reasoning features are highly sensitive to token-level interventions. Injecting a small number of feature-associated tokens into non-reasoning text is sufficient to elicit strong activation for 59% to 94% of features, indicating reliance on lexical artifacts. For the remaining features that are not explained by simple token triggers, LLM-guided falsification consistently produces non-reasoning inputs that activate the feature and reasoning inputs that do not, with no analyzed feature satisfying our criteria for genuine reasoning behavior. Steering these features yields minimal changes or slight degradations in benchmark performance. Together, these results suggest that SAE features identified by contrastive approaches primarily capture linguistic correlates of reasoning rather than the underlying reasoning computations themselves.

Figures (9)

• Figure 1: Overview of our falsification-based evaluation framework for reasoning features. We begin with SAE features identified by contrastive activation differences. We then test whether lexical cues are sufficient to induce activation via causal token injection. Remaining features are subjected to LLM-guided adversarial counterexample generation, which decouples reasoning behavior from feature activation. Across all configurations, this process fails to identify any feature that satisfies our criteria for genuine reasoning behavior.
• Figure 2: Token concentration ratio and normalized activation entropy for SAE features across all layers of Gemma-3-12B-Instruct on the s1K dataset. Middle layers exhibit lower concentration and higher entropy, indicating reduced reliance on specific tokens.
• Figure 3: Distribution of Cohen's $d$ values across SAE features for Gemma-3-12B-Instruct at layer 22.
• Figure 4: Distribution of token injection classifications across configurations. Each bar corresponds to a model and layer configuration, with segments indicating the proportion of features classified as token-driven (TD), partially TD, weakly TD, or context-dependent.
• Figure 5: Jaccard similarity between top-100 feature sets selected by different ranking metrics on Gemma-3-4B-Instruct at layer 22 using the s1K dataset.