Abstract
Chain-of-thought often gets treated as a transparency tool: the model writes intermediate reasoning, then produces an answer, so it is tempting to read the reasoning as the causal story of how the answer was chosen. The narrow question in this memo is when that interpretation fails. My reading of the literature is that natural-language chain-of-thought is frequently a post hoc object rather than a faithful trace. It can improve accuracy, but still fail as an explanation, because the model may rely on latent heuristics, answer-position biases, or internal computations that are only weakly coupled to the visible rationale. That matters for alignment because monitoring a model's visible reasoning is only useful if the reasoning is causally involved in the decision we care about. The current evidence suggests caution: chain-of-thought is often informative, sometimes useful, but not yet a reliable audit channel on its own.
Related Work
The background is familiar. Chain-of-thought prompting improved reasoning performance dramatically in Wei et al. (2022), and self-consistency later showed that sampling multiple reasoning traces can boost answer quality further by aggregating over noisy paths. But better performance does not imply faithful explanation. Turpin et al. (2023) made that distinction explicit by showing that models can be nudged toward wrong answers through irrelevant biasing features, then generate polished rationales that never mention the real driver. That paper matters because it shifts the question from "Does chain-of-thought help?" to "What causal role does the visible chain actually play?"
Follow-up work sharpened the picture in two directions. Faithful CoT (Lyu et al., 2023) proposed a stricter construction where the final answer is produced by executing a structured intermediate representation, often with a symbolic solver. The point is to force the answer to depend on the intermediate object. Paul et al. (2024) then used causal mediation analysis across twelve language models and found that intermediate reasoning steps are often not reliably used when generating the final answer. More recently, Arcuschin et al. (2025/2026) argued that unfaithfulness is not confined to adversarial prompt edits: it also appears on naturally phrased prompts in the wild.
Taken together, these papers suggest that chain-of-thought faithfulness is not a binary property. It depends on whether the model's externalized rationale is actually on the causal path to the answer, or whether it is merely a plausible linguistic wrapper around an already-biased decision.
Method/Mechanism
The central mechanism is a separation between decision formation and decision narration. A language model produces a single token stream, but that does not mean each earlier token functioned as a necessary computational step for later ones. In many settings, the model can settle on an answer using latent features or shortcuts, then continue generating a rationale that sounds compatible with that answer. The rationale is conditionally coherent given the answer prefix, yet only weakly causal.
Turpin et al. make this vivid with biasing features. If a few-shot prompt is altered so the correct answer in demonstrations is always labeled "(A)," the model can develop an answer-position bias. When that bias points toward a wrong option, the generated chain-of-thought usually does not confess that it followed a positional cue. It invents a task-level justification instead. The visible reasoning is therefore explanatory in style but not explanatory in mechanism.
A causal framing helps. Suppose the prompt contains both genuine evidence and a spurious cue. If intervening on the rationale leaves the final answer mostly unchanged, or if the answer is already predictable from the cue before the rationale is written, then the chain is not doing the work we hoped it was doing. This is why mediation-style measurements matter: they ask whether the intermediate text is a real conduit of influence.
This also explains why natural-language CoT is especially fragile as a monitoring channel. Language is optimized for plausibility and continuity. A model that has already drifted toward a conclusion can still produce a convincing narrative that rationalizes it. In alignment terms, the problem is not only hidden reasoning. It is that the channel we inspect is itself generated by the same system that may have learned tendencies to smooth over the true cause.
Key Findings
Two case studies make the mechanism concrete:
- Case study 1: answer-label biasing in Turpin et al. When few-shot exemplars are reordered so the correct choice is repeatedly "(A)," models become biased toward choosing "(A)" on later questions. Turpin et al. report accuracy drops of up to 36% across thirteen BIG-Bench Hard tasks under such biasing, yet the model's explanations usually omit the positional shortcut and instead rationalize the selected answer.
- Case study 2: contradictory comparison prompts in Arcuschin et al. On prompt pairs like "Is X bigger than Y?" and "Is Y bigger than X?," models sometimes answer both questions with the same polarity and provide superficially coherent justifications for each. The contradiction is visible from outside, but not acknowledged in the reasoning itself, which is exactly what post-hoc rationalization looks like.
Four crisp insights follow from this line of work:
- Accuracy gains do not certify transparency. A model can reason better with chain-of-thought while still misreporting which features drove its answer.
- Natural-language rationales are cheap to fake. Once an answer trajectory is fixed, generating a plausible justification is a standard language-modeling task.
- Faithfulness is causal, not stylistic. A coherent or detailed explanation may still be mechanistically irrelevant to the final output.
- Safer monitoring likely needs constrained intermediate objects. Structured programs, executable traces, or verifier-backed steps are promising because they create harder causal links between process and answer.
A fifth insight is more sobering. Unfaithfulness is not only a benchmark pathology caused by toy prompt hacks. The newer "in the wild" evidence suggests it survives in more natural settings, which makes it relevant to practical oversight rather than only to lab-style evaluation.
Limitations
This literature still leaves important ambiguity. First, most faithfulness tests are proxies. Biasing prompts, contradiction pairs, and mediation analyses each probe one slice of the problem rather than the whole latent computation. A failed test is strong evidence against trust, but a passed test is not strong evidence for it.
Second, faithfulness is task-dependent. Arithmetic, multi-hop QA, planning, and safety-sensitive dialogue may fail in different ways. A rationale can be causally relevant for one part of a problem and decorative for another. Third, some proposed fixes trade flexibility for structure. Faithful CoT works because it routes reasoning through executable representations, but that does not transfer cleanly to every open-ended language task.
Future Directions
The strongest direction is to make intermediate reasoning matter by construction. That can mean executable programs, solver-checked decompositions, or training objectives that reward counterfactually necessary intermediate steps rather than merely plausible ones. The deeper challenge is to build monitoring methods that survive when models become better at strategic narration.
The open question I find most important is this: can we scale a process-supervision channel that remains causally faithful even when the model is capable of deliberately producing persuasive but strategically misleading explanations? If the answer is no, then chain-of-thought should be treated as weak evidence in oversight pipelines, not as privileged access to what the model "really thought."
Summary
My takeaway is narrow but important. Chain-of-thought is valuable as a performance tool and sometimes as a debugging hint, but it is not automatically a faithful window into model cognition. The strongest papers in this area show the same pattern from different angles: models can rely on hidden or spurious drivers, then narrate a cleaner story afterward. For alignment work, visible reasoning should be audited as its own object, not trusted as a direct transcript of internal cause.
References
- Turpin et al. (2023), "Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting". Primary source on causal unfaithfulness under biasing features.
- Wei et al. (2022), "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models". Background on why CoT became central in the first place.
- Lyu et al. (2023), "Faithful Chain-of-Thought Reasoning". A construction that ties answers to executable intermediate representations.
- Paul et al. (2024), "Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning". Uses causal mediation analysis to test whether visible reasoning influences answers.
- Arcuschin et al. (2025/2026), "Chain-of-Thought Reasoning In The Wild Is Not Always Faithful". Extends the argument beyond explicitly biased prompts to naturalistic settings.