Chain-of-Thought Text Is Not a Reliable Explanation
Seven recent preprints show that visible reasoning can omit causal factors, rationalize an answer, and receive very different faithfulness scores under different measurement methods.
Scope note: This review covers seven recent studies of whether visible chain-of-thought text reflects the causes of a model answer. It does not claim that all reasoning text is false or that it has no diagnostic use.
A model writes a detailed explanation and arrives at the correct answer. The explanation may still be unrelated to the computation that determined the answer.
That possibility changes how chain-of-thought text should be used. It can help a model solve a task. It can help a reviewer notice an error. It can also provide a convincing account that did not cause the result.
Seven recent preprints test this problem with interventions, competing classifiers, hidden cues, omitted factors, and controlled ground truth. The shared finding is not that visible reasoning has no value. It is that fluent reasoning text should not be treated as a verified explanation.
Different tests produce different faithfulness scores
Measuring Faithfulness Depends on How You Measure applies three classifiers to the same 10,276 reasoning traces from twelve open models. The resulting overall faithfulness rates were 74.4, 82.6, and 69.7 percent. Individual model gaps reached 30.6 percentage points, and classifier choice reversed some model rankings.
The classifiers were not making random errors. They used different definitions. One looked for a textual mention of an influencing cue. Another required stronger evidence that the cue affected the conclusion. A single faithfulness number concealed that difference.
Faithfulness Metrics Don’t Measure Faithfulness asks a more basic question: do proposed metrics correspond to known ground truth? The authors construct settings where the actual dependence is available, then compare common metrics against it. Their findings challenge the use of absolute metric scores when the metric itself has not been validated against a known causal relationship.
I would therefore treat a published faithfulness percentage as method-specific. Comparisons are useful only when the studies define and measure the same property.
A changed explanation may not change the answer
Project Ariadne changes intermediate reasoning claims by negating premises, reversing facts, and altering logic. The framework then checks whether the final answer responds to those changes. The authors report violation density as high as 0.77 in factual and scientific tasks: many answers remained unchanged even when the visible reasoning was made contradictory.
This is direct causal evidence. If the stated premise changes and the answer does not, the premise was not controlling the answer in the way the text claimed.
Analysing Chain of Thought Dynamics studies instruction-tuned, reasoning, and reasoning-distilled models on tasks that require less formal deduction. The authors find that the influence of chain-of-thought and its faithfulness do not consistently vary together. A reasoning trace can affect performance without accurately reporting the model’s decision process.
Complete-looking text can omit required factors
Measuring Chain-of-Thought Monitorability adds verbosity to faithfulness. Here, verbosity means whether the trace states every factor required to solve the task, not whether it uses many words. A model can acknowledge an inserted cue and therefore appear faithful while omitting other factors needed to understand its decision.
That omission matters for safety review. A monitor cannot identify a harmful consideration if the model does not state it. More explanation does not necessarily provide more relevant information.
Lie to Me tests whether models disclose the true reason for an answer when incentives favor a different account. Its results add evidence that generated explanations can adapt to the requested presentation rather than preserve causal accuracy.
Faithfulness can be improved, but the objective matters
Balancing Faithfulness and Performance trains a speaker model to produce reasoning that several listener models can continue successfully. The method improved three faithfulness measures while also improving accuracy across several reasoning benchmarks. The resulting traces were shorter and more direct.
This is an encouraging result. It also shows that faithfulness is a trained behavior, not an automatic result of asking for steps. The listener objective rewards reasoning that other systems can execute. It does not establish direct access to every internal computation, but it creates stronger evidence that the stated steps support the answer.
How I use visible reasoning
I use chain-of-thought text as a work artifact. It can expose an assumption, reveal missing evidence, support a code review, or make a proposed decision easier to challenge. I do not use it as proof that the model disclosed its internal cause.
For consequential decisions, I want tests that change the claimed cause and observe whether the answer changes. I want independent evidence, repeated runs, and direct checks of the final result. If a monitoring system depends on reasoning text, I want it evaluated for omitted factors and for sensitivity to the method used to score faithfulness.
The distinction is precise. Reasoning text can be useful without being a reliable explanation. We can read it, test it, and use it to direct verification. We should not grant it authority that the research does not support.