The Google DeepMind research team has recently achieved a significant breakthrough with the development of an innovative technology called SCoRe (Self-Correction through Reinforcement Learning). This technology aims to address the long-standing challenge of large language models (LLMs) struggling to self-correct, enabling them to identify and fix errors without relying on multiple models or external checks. The core of SCoRe's technology lies in its two-phase approach. The first phase optimizes the model.

Google DeepMind's latest research finds that large language models have difficulty autonomously correcting reasoning errors. The study found that models often err when trying to correct their initial responses solely based on their own judgment. The researchers believe that in currently sensitive safety scenarios, self-correction capabilities are essential. Achieving consensus responses through voting can ensure consistency but is not equivalent to self-correction. Further research is needed to improve the self-correction abilities of language models.