New AI Technique Provides Step-by-Step Feedback to Reduce Errors
-
Chain-of-feedback (CoF) is a new prompting technique to steer generative AI away from errors and hallucinations. It provides step-by-step guidance during the AI's response process.
-
CoF is useful when AI gives incorrect answers. Rather than endlessly prompting the AI to "try again", CoF breaks the problem into steps and provides feedback on each step.
-
Recursive CoF takes problematic steps, asks a separate AI to solve just that step, then feeds the corrected step back into the original problem. This further guides the AI.
-
Best practices for CoF include avoiding "mindless" redos, inspecting stepwise responses to catch errors early, deriving corrections, and repeating the process until satisfied or progress stalls.
-
CoF should be used judiciously on hard problems, not applied universally. Practice helps gain facility with this technique.