Self-Refine
A training-free self-reflection method that iteratively improves model outputs through a generate-critique-revise loop without updating model parameters.
Core Facts
Timeline (last 90 days)
Training-free self-reflection protocols refer to methods that do not update model parameters and instead let the model self-check and revise at inference time through prompting or workflow design, typically represented by the "generate→critique→revise" loop (such as Self-Refine).
Existing training-free self-reflection protocols (generate→critique→revise) operate on flat token streams, so revisions cannot locate which specific step is unreliable, and early stopping cannot exploit structural convergence signals.
Ideas similar to generate-verify-regenerate also appear in academic methods such as Self-Refine and Reflexion.
Methods such as CRITIC, Self-Refine, and Reflexion have emerged in the field of self-correction, and CriticGen can be viewed as the latest development in this technical lineage.
The core mechanism of Self-Refine is to have the large language model critically evaluate its own initial output, then generate an improved version based on the evaluative feedback, iterating until requirements are met or the iteration limit is reached, with the entire process relying solely on a single model's reasoning capability and requiring no external reward model or human feedback.
Self-Refine is an iterative self-optimization framework requiring no additional training, proposed in 2023 by Carnegie Mellon University and other institutions.
The direction explored by Ornith-1.5 echoes self-reflection frameworks proposed in academia such as Self-Refine and Reflexion.
All Facts (9)
The Reflexion or Self-Refine mechanism was systematically proposed by Shinn et al. in a 2023 paper.
65%Unverified多角色流水线（Self-Refine）的输出质量通常比单次LLM调用提升15%-30%
60%UnverifiedTraining-free self-reflection protocols refer to methods that do not update model parameters and instead let the model self-check and revise at inference time through prompting or workflow design, typically represented by the "generate→critique→revise" loop (such as Self-Refine).
50%UnverifiedExisting training-free self-reflection protocols (generate→critique→revise) operate on flat token streams, so revisions cannot locate which specific step is unreliable, and early stopping cannot exploit structural convergence signals.
50%UnverifiedIdeas similar to generate-verify-regenerate also appear in academic methods such as Self-Refine and Reflexion.
50%UnverifiedMethods such as CRITIC, Self-Refine, and Reflexion have emerged in the field of self-correction, and CriticGen can be viewed as the latest development in this technical lineage.
50%UnverifiedThe core mechanism of Self-Refine is to have the large language model critically evaluate its own initial output, then generate an improved version based on the evaluative feedback, iterating until requirements are met or the iteration limit is reached, with the entire process relying solely on a single model's reasoning capability and requiring no external reward model or human feedback.
50%UnverifiedSelf-Refine is an iterative self-optimization framework requiring no additional training, proposed in 2023 by Carnegie Mellon University and other institutions.
50%UnverifiedThe direction explored by Ornith-1.5 echoes self-reflection frameworks proposed in academia such as Self-Refine and Reflexion.
50%