Self-Verification
Self-verification is the process where the same LLM that generated an answer critiques and verifies it, re-deriving the solution or checking each step for consistency without an external oracle.
What is Self-Verification?
Self-verification is the process where the same LLM that generated an answer critiques and verifies it, re-deriving the solution or checking each step for consistency without an external oracle.
Self-verification is the process where the same LLM that generated an answer critiques and verifies it, re-deriving the solution or checking each step for consistency without an external oracle.
Where is it used?
Self-Refine and Self-CheckGPT use the model to verify its own outputs; DeepSeek-R1's reasoning trace includes self-correction steps; this improves math and code accuracy.
How to build it
After generating a solution, prompt `"Review your answer step by step. If an error is found, correct it."` in a second LLM call, and return the revised answer; compare accuracy with and without.