New RGR technique mitigates LLM capability degradation during self-improvement

A new research paper introduces Representation Geometry Regularization (RGR) to address capability degradation in large language models (LLMs) during self-improvement processes like on-policy self-distillation. The study found that while LLMs can improve overall performance through self-distillation, they may lose previously acquired capabilities. RGR aims to preserve representation geometry by promoting compactness within clusters and separation between clusters, integrating seamlessly with existing methods. Experiments indicate RGR effectively mitigates this degradation while maintaining performance gains. AI

IMPACT This research could lead to more robust LLM self-improvement techniques, reducing the risk of losing valuable capabilities during training.

RANK_REASON Academic paper detailing a new technique for LLM self-improvement. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• large language models
• On-Policy Self-distillation
• Representation Geometry Regularization

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →