Self-correction
An agent capability where the model evaluates its own output for errors and produces a corrected version — improves accuracy on verifiable tasks by 10–30% in 2026.
Self-correction is the pattern where an agent produces an answer, checks it, finds the error, and fixes it. The simplest implementation: emit an answer, then prompt the same model with "review this for errors and produce a corrected version." More sophisticated versions use separate critic models or explicit verification tools (run tests, check facts).
In 2026 most production agents implement some form of self-correction. Coding agents run tests and fix failures. Math agents verify with structured checks. Customer support agents review their drafts before sending. The pattern adds cost but lifts accuracy meaningfully on verifiable tasks.
The limit is the model itself. If a model cannot detect its own error, self-correction fails. For the hardest problems, you need an outside verifier — tests, structured checks, or a stronger model in the critic role.
Frequently asked
Does self-correction always improve quality?+
Only when the model can detect its own errors. On hard reasoning tasks where the model is overconfident, self-correction can entrench the wrong answer. Pair self-correction with verifiable checks (tests, structured validation) for best results.
Self-correction vs reflexion — what is the difference?+
Self-correction happens within a single attempt: produce, critique, correct. Reflexion happens across multiple attempts: try, reflect, try again. Both improve accuracy; reflexion is more powerful but more expensive.