Skip to content
Go back

Chain-of-Thought Prompting Can Improve Answers; It Is Not Proof

Updated:
Commentary

Originally published onLinkedIn on 12 Jul 2026.

My original LinkedIn post shared Amit Shekhar’s accessible explanation of chain-of-thought prompting. Shekhar deserves credit for making the basic mechanism and its common variants approachable. My short summary repeated the popular formulation—ask a model to reason step by step, let it “show its work,” and accuracy improves—but compressed two different claims into one.

The evidence supports the narrower claim that eliciting intermediate steps can improve measured performance for some models, tasks, prompts, and decoding strategies. It does not establish that a visible rationale is a faithful record of the process that produced the answer. Nor does a fluent sequence of steps certify that the conclusion is correct.

That distinction matters anywhere an AI answer is expected to survive verification rather than merely look convincing.

What the original results showed

Wei et al. demonstrated few-shot chain-of-thought prompting: a prompt contains worked examples with intermediate reasoning, and the model continues that pattern for a new problem. They reported substantial gains on selected arithmetic, commonsense, and symbolic reasoning benchmarks, especially with sufficiently large models.

Kojima et al. studied a different method. Their zero-shot approach appended an instruction such as “Let’s think step by step” without providing worked examples. It also improved results on several reasoning benchmarks used in that study.

These are important empirical results, but neither paper supports the universal rule “more visible reasoning means a more accurate answer.” Performance depends on the task, model, prompt, and evaluation. A technique that helps on a multi-step arithmetic benchmark may add latency or failure opportunities to a simple classification or retrieval task.

Several techniques share the same label

The term chain of thought now covers practices with materially different mechanisms and evidence.

TechniqueWhat changesWhat it does not guarantee
Few-shot CoTWorked reasoning examples are included in the promptThat the examples induce the right method
Zero-shot CoTA short instruction elicits an intermediate explanationThat the instruction improves every task
Self-consistencySeveral reasoning paths are sampled and aggregatedThat a majority answer is correct or independent
Reasoning modelsAdditional inference is performed before the final replyThat the public explanation is an internal trace

Self-consistency is particularly revealing. Instead of trusting one greedy reasoning path, it samples multiple paths and selects the most consistent answer. Its gains show that decoding and aggregation matter, not only whether the model emits a step-by-step explanation. Agreement can still be wrong, but the technique changes the evidence from one plausible story to a distribution of attempted solutions.

A plausible explanation is not evidence

The temptation is to read generated reasoning as an audit log. That interpretation is unsafe.

Turpin et al. showed that biasing features could affect model predictions while the generated explanations systematically failed to mention those influences. When a bias pushed a model toward an incorrect answer, the model could produce a coherent rationale for that answer after the fact.

Later work by Arcuschin et al. found unfaithful reasoning on naturally worded prompts as well, including cases where models constructed superficially coherent arguments for contradictory answers. The rates and behaviours vary by model and setup, but the engineering conclusion is stable: verbalized reasoning is not a complete account of the process that generated the result.

A chain of thought may expose a useful intermediate claim. It may also rationalize a prior answer, omit a decisive influence, contain a subtle inconsistency, or arrive at the correct result through invalid steps. Fluency makes those cases harder—not easier—to distinguish by inspection alone.

Verification changes the question

The operational question should not be “Did the model explain itself step by step?” It should be “Which claims can be checked independently?”

The appropriate checks depend on the domain:

In each case, the verifier should be at least partly independent of the prose being evaluated. Asking the same model to declare its own reasoning correct is additional testimony, not independent evidence.

A practical prompting and testing pattern

For high-consequence work, a more defensible workflow is:

  1. Ask for the result, assumptions, and a concise decomposition into checkable claims.
  2. Request multiple candidate solutions when the problem admits meaningfully different approaches.
  3. Send calculations, code, citations, or proposed actions to an external verifier.
  4. Preserve verifier output and observed effects as evidence.
  5. Test boundary cases and actively try to falsify the answer.
  6. Report remaining uncertainty instead of filling it with a longer rationale.

This pattern still benefits from decomposition. It simply refuses to confuse the decomposition with proof. The generated steps become hypotheses and test targets rather than privileged access to the model’s internal process.

When chain of thought is useful

Chain-of-thought prompting remains useful when decomposition helps a model solve a multi-step task, when intermediate claims guide debugging, or when those claims can be checked with tools. It is less attractive for simple tasks, for settings where verbose output increases cost or leaks sensitive context, and for problems with a cheap authoritative verifier that can evaluate the answer directly.

The corrected conclusion is therefore narrower than my original LinkedIn post, but more useful: chain of thought is an inference technique and a diagnostic interface. It can improve answers under particular conditions and expose material worth testing. It is not a correctness certificate, a proof, or a guaranteed faithful transcript of how a model arrived at its answer.


Share this note

Previous Note
Property-Based Testing Starts with the Property, Not the Generator
Next Note
Agentic Resource Discovery Still Needs a Verification Boundary