TL;DR
The capability of an AI model to autonomously detect, evaluate, and resolve its own errors during the inference or generation process without human intervention.
Self-correction involves multi-turn processes where a large language model generates an initial response, evaluates it for logical, factual, or grammatical errors, and systematically revises the output. While traditional models struggle to correct themselves without external prompts, modern approaches train models using reinforcement learning to generate internal reasoning traces that natively catch and fix coding or math mistakes. This eliminates reliance on separate teacher models or secondary prompt iterations.
Why this matters for your business
It enables AI agents to work autonomously on complex, multi-step engineering and reasoning tasks by verifying and fixing their own work before presenting the final result.
