Reflective Recovery
A self-supervised fine-tuning method for enhancing the reasoning capabilities of large language models. It extracts initial segments from failed reasoning trajectories as training data, enabling the model to learn to identify and correct errors during reasoning without relying on external judge models or reinforcement learning frameworks.
Timeline (last 90 days)
An arXiv paper (arXiv:2609.19156) proposes the Reflective Recovery method.
Reflective Recovery is a self-supervised method that converts the model's failed reasoning attempts into recovery training data.
Using Reflective Recovery, DeepSeek-R1-Distill-Qwen-7B's accuracy on AIME 2025 improved from 30.0% to 37.5%.
The Reflective Recovery method breaks through the barrier of Scaling Collapse.
Models trained with Reflective Recovery exhibit emergent self-correction behavior.
The paper's authors summarize this transition as a shift from outcome-oriented memorization to process-oriented reflective reasoning.
The approach of Reflective Recovery is analogous in the reinforcement learning domain to off-policy learning using failed trajectories.
This work is currently still an academic preprint, and its generalization ability to larger-scale models and broader tasks remains to be verified by subsequent research.
All Facts (8)
This work is currently still an academic preprint, and its generalization ability to larger-scale models and broader tasks remains to be verified by subsequent research.
50%UnverifiedReflective Recovery is a self-supervised method that converts the model's failed reasoning attempts into recovery training data.
50%UnverifiedUsing Reflective Recovery, DeepSeek-R1-Distill-Qwen-7B's accuracy on AIME 2025 improved from 30.0% to 37.5%.
50%UnverifiedThe Reflective Recovery method breaks through the barrier of Scaling Collapse.
50%UnverifiedModels trained with Reflective Recovery exhibit emergent self-correction behavior.
50%UnverifiedThe paper's authors summarize this transition as a shift from outcome-oriented memorization to process-oriented reflective reasoning.
50%UnverifiedThe approach of Reflective Recovery is analogous in the reinforcement learning domain to off-policy learning using failed trajectories.
50%UnverifiedAn arXiv paper (arXiv:2609.19156) proposes the Reflective Recovery method.
50%