Do Sparse Autoencoders Identify Reasoning Features in Language Models?

Problem Statement
Recent advances in large language models (LLMs) have demonstrated strong multi-step reasoning capabilities, often elicited via chain-of-thought (CoT) prompting. A growing body of mechanistic interpretability research utilizes Sparse Autoencoders (SAEs) to identify internal features associated with these reasoning processes. The prevailing approach involves contrastive selection: identifying SAE features that exhibit significantly higher activation on reasoning text (e.g., CoT traces) compared to non-reasoning text. However, the authors argue that this method suffers from fundamental ambiguities. Because CoT responses differ from direct answers not only in underlying reasoning but also in surface lexical usage (e.g., discourse markers like "Wait," "But," "Let"), contrastive methods may identify features that capture shallow lexical confounds rather than genuine reasoning computations. The central question addressed is whether SAEs can reliably identify monosemantic features that represent the internal reasoning process itself, or if they merely capture low-dimensional linguistic correlates that co-occur with reasoning.

Methodology
The paper proposes a falsification-based evaluation framework designed to test the robustness of contrastively selected "reasoning features." The methodology proceeds in three stages:

1. Theoretical Motivation: The authors present a stylized analysis showing that sparsity-regularized decoding in SAEs inherently favors stable, low-dimensional correlates over high-dimensional within-behavior variation. In a setting where a high-level behavior (reasoning) co-occurs with a stable lexical cue, sparse decoding can preferentially retain the cue coordinate while suppressing the distributed, high-dimensional variation of the reasoning trace itself. This suggests that contrastive methods may isolate cue-like structures rather than reasoning computations.

2. Causal Token Injection: To test for superficial lexical dependencies, the authors perform causal token injection experiments. They identify the top-activating tokens or n-grams for a candidate feature and insert them into non-reasoning text. If the feature activates significantly upon the injection of these few tokens, it is classified as "token-driven," indicating it captures a lexical cue rather than a reasoning process.

3. LLM-Guided Falsification: For features that remain "context-dependent" after token injection, the authors employ an LLM-guided adversarial process. An LLM generates hypotheses about what linguistic or structural patterns a feature detects. It then constructs two types of counterexamples:

   • False Positives (FP): Non-reasoning text that instantiates the hypothesized pattern (e.g., specific discourse markers or syntactic structures) to trigger activation without reasoning.
   • False Negatives (FN): Paraphrases of top-activating reasoning traces that preserve the semantic meaning and reasoning intent but alter the surface form (vocabulary, syntax, style) to suppress activation.
     A feature is considered a "genuine reasoning feature" only if it resists these falsification attempts (i.e., it does not activate on FPs and does activate on FNs).

4. Steering Sanity Check: As a supplementary check, the authors perform feature steering experiments to see if amplifying candidate features improves downstream reasoning benchmarks, noting that performance changes alone are insufficient proof of reasoning representation.

Key Contributions

• Theoretical Insight: The paper provides a theoretical mechanism explaining why sparsity in SAEs creates an asymmetry that favors the retention of stable, low-dimensional lexical cues over high-dimensional reasoning variability, even when the dictionary is perfect.
• Evaluation Framework: It introduces a rigorous, falsification-oriented pipeline combining causal token injection and LLM-guided counterexample generation to distinguish between genuine reasoning features and linguistic correlates.
• Empirical Evidence: The study systematically applies this framework across 22 configurations involving multiple model families (Gemma-3, DeepSeek-R1-Distill-Llama, Llama-3.1), layers, and reasoning datasets.

Results

• Token Sensitivity: Across all configurations, a significant majority (45%–90%) of contrastively selected candidate features were found to be highly sensitive to token-level interventions. Injecting only a few associated tokens into non-reasoning text was sufficient to trigger substantial activation increases, classifying them as token-driven or partially token-driven.
• Falsification of Context-Dependent Features: For the remaining features that were not explained by simple token injection, the LLM-guided falsification process successfully constructed targeted counterexamples. The LLM generated non-reasoning texts that activated the features (False Positives) and paraphrased reasoning texts that suppressed activation (False Negatives).
• Zero Genuine Features: Across all 248 context-dependent features analyzed, zero were classified as genuine reasoning features. Every feature examined was shown to be a confound, typically detecting specific syntactic templates (e.g., "First, I need to..."), discourse markers, or formal academic framing rather than the underlying reasoning computation.
• Steering Impact: Steering experiments yielded minimal or slightly degraded performance on benchmarks (AIME, GPQA), consistent with the finding that these features do not encode the reasoning mechanism itself.

Significance and Claims
The paper concludes that in the settings studied, sparse decompositions tend to favor low-dimensional correlates that co-occur with reasoning rather than isolating the reasoning process itself. The authors claim that contrastive correlations alone are insufficient evidence for the existence of monosemantic reasoning features. They emphasize that attributing high-level behaviors to individual SAE features requires causal validation and falsification to rule out superficial linguistic confounds. The work does not claim that SAEs cannot represent reasoning-relevant structure, but rather that current contrastive pipelines reliably identify features that are better explained as linguistic correlates. The authors advocate for treating contrastive correlations as hypotheses that must undergo rigorous causal and adversarial testing before being interpreted as representations of specific cognitive computations.