Bypassing the Rationale: Causal Auditing of Implicit Reasoning in Language Models

TL;DR

This paper introduces a causal layerwise auditing method for chain-of-thought prompting in language models, revealing that reasoning influence is often localized and can be bypassed, challenging assumptions about transparency.

Contribution

It presents a novel activation patching approach and the CoT Mediation Index to measure causal influence of reasoning, uncovering variability across models and tasks.

Findings

CoT influence is often localized in narrow reasoning windows.

Models tuned for reasoning show stronger, more structured mediation.

Bypass regimes exist where reasoning text has minimal causal impact.

Abstract

Chain-of-thought (CoT) prompting is widely used as a reasoning aid and is often treated as a transparency mechanism. Yet behavioral gains under CoT do not imply that the model's internal computation causally depends on the emitted reasoning text, i.e., models may produce fluent rationales while routing decision-critical computation through latent pathways. We introduce a causal, layerwise audit of CoT faithfulness based on activation patching. Our key metric, the CoT Mediation Index (CMI), isolates CoT-specific causal influence by comparing performance degradation from patching CoT-token hidden states against matched control patches. Across multiple model families (Phi, Qwen, DialoGPT) and scales, we find that CoT-specific influence is typically depth-localized into narrow "reasoning windows," and we identify bypass regimes where CMI is near-zero despite plausible CoT text. We further observe that models tuned explicitly for reasoning tend to exhibit stronger and more structured mediation than larger untuned counterparts, while Mixture-of-Experts models show more distributed mediation consistent with routing-based computation. Overall, our results show that CoT faithfulness varies substantially across models and tasks and cannot be inferred from behavior alone, motivating causal, layerwise audits when using CoT as a transparency signal.