Mid · 7 chapters
Reflection and self-critique in a single prompt, and its limits
Models are poor at spotting their own errors unaided
Self-critique is useful only when treated as an unreliable component whose gains are measured against regression, latency, and token cost.
Read these first
What you will know
Add a self-critique pass and measure whether it actually helped
Use a fixed evaluation set and report the delta in accuracy, error types, latency, and token usage rather than trusting more polished answers.
Show a case where the model fails to spot its own error
Models are poor at spotting their own errors, because the same weights that made the mistake are judging it.
Reproduce a case where repeated unaided critique stops improving the answer
A model can reinforce a false premise or over-edit a correct answer while producing increasingly confident justifications.
Implement a separate critic prompt or objective verifier and compare it with single-pass output (11.5–11.7)
An independent signal such as a schema check, test, retrieval comparison, or second prompt makes the revision decision auditable.
Stop a critique loop that has begun making things worse
Use an iteration cap and regression gate so a revised answer is retained only when it beats the previous version on the target checks.
Chapters
- 1
Unaided self-critique
Asking a model to check its own answer runs a consistency check against itself, and the answer is consistent with itself by construction: a summary that dates a merger to March 2021 when the source says 2019 is reported as accurate, because the same weights that generated the date are judging it. Covers why an appraisal-shaped review turn comes back clean, what a lookup-shaped check finds instead, and what does and does not change when the review moves to a separate call with no prior reasoning in context.
- 2
The checker independence ladder
What a checker is worth is decided by how correlated its errors are with the generator's, not by how accurate it is on its own: a critic that is 88% accurate overall and 15% accurate on the cases the generator got wrong is telling you its blind spots are the generator's blind spots. Covers the four rungs — same context, fresh context, a different model family, and a verdict that comes from no model at all — what each catches and misses, and the two numbers that belong in a design doc instead of overall accuracy.
- 3
Evaluating a critique pass
A second turn that moves execution-match accuracy from 71% to 72% on 200 cases, doubles p95 and raises output tokens 2.4x has produced three numbers that cannot support any conclusion, because one point on 200 cases is two cases and the baseline moves 3.5 points on its own. Covers the paired 2x2 of flips, the exact test on the discordant cases, severity weighting so a loud fix and a silent regression stop counting the same, and cost per net corrected case.
- 4
Answer quality across reflection rounds
Loop critique and revision five times and accuracy peaks at round two, then falls while the answers get longer, lose their hedges and acquire invented flags with defaults and migration procedures. Two mechanisms compound: a round asked to find fault manufactures one on an already-correct answer, and each round conditions on the last round's justifications, so a mistaken premise gets restated as settled. Covers the one-afternoon sweep that finds where a loop peaks, and why a model judge reports monotone improvement from a loop that is degrading.
- 5
Objective verification signals
Schema validation, a total recomputed from line items and a provenance substring test are checks with stateable coverage, and using the same signal twice — as the trigger for a revision and as the acceptance test on it — keeps the model's opinion of its own work out of the decision entirely. Covers what five of six invoice error classes cost in deterministic code, the class nothing covers, and the job a critic prompt can still hold once adjudication has been taken away from it.
- 6
Iteration caps and regression gates
A loop that returns whatever the last round produced will hand back an elaboration of an answer that was already correct at round one, which on a 240-case eval was 23 cases and 6.7 points. Covers splitting the round counter into two independent pieces — an iteration cap that bounds the degrading region and a gate that keeps a revision only when it beats the incumbent on hard checks and a soft score by a margin set from the metric's own noise — plus what to do when a model has to break a tie.
- 7
Routing requests into a critique pass
A critique pass that doubles p95 and takes the monthly bill from $3,530 to $7,410 is being applied to 420,000 requests to repair errors in a few tens of thousands of them. Covers building the trigger from the paired eval's own labels so it predicts critique-fixable errors rather than errors, the four-configuration comparison that keeps 91% of the gain for 14% of the incremental spend, and the shadow sample that tells you what the router has started missing.
External resources
- Large Language Models Cannot Self-Correct Reasoning Yetarxiv.org
The core result: unaided self-critique often makes answers worse.
- Self-Critique & Self-Refine: LLMs Checking Their Work | AI/TLDRai-tldr.dev
This is a substantive explainer with concrete generate–critique–refine mechanics, prompt guidance, and clear discussion of self-correction's failure modes and limits.
- Feedback Friction: LLMs Struggle to Fully Incorporate External Feedbackarxiv.org
Even correct external feedback is only partially absorbed; plan for it.
- CRITICTOOL: Evaluating Selfaclanthology.org
Evaluates whether single-prompt self-critique helps models detect errors—and where it fails.
- AI Agents 4 - Prompt Chaining, Routing and ReflectionProf. Ghassemi Lectures and Tutorials
Implements the reflection/self-critique pattern and where it stops helping