{"slug": "mitigating-capability-degradation-in-llm-self-improvement-via-representation", "title": "Mitigating Capability Degradation in LLM Self-Improvement via Representation Geometry Regularization", "summary": "A new arXiv paper (2610.12855v1) reports that On-Policy Self-distillation (OPSD) can progressively degrade previously acquired capabilities in large language models, with queries answered correctly at earlier training stages becoming incorrect as optimization proceeds even while overall performance improves. The authors attribute the degradation to deteriorating hidden-representation geometry — reduced intra-cluster compactness and diminished inter-cluster separation — and propose Representation Geometry Regularization (RGR), which preserves that geometry during post-training and can be integrated into existing OPSD methods without altering their core optimization procedures. Experiments across multiple OPSD methods and benchmarks show RGR generally alleviates capability degradation while retaining capability improvements, with code released at github.com/supermachinelearner/RGR.", "body_md": "arXiv:2610.12855v1 Announce Type: new \nAbstract: On-Policy Self-distillation is a self-improvement paradigm that enables large language models (LLMs) to improve their capabilities by leveraging the same model as both teacher and student under different conditioning contexts. However, we find that OPSD methods can progressively degrade previously acquired capabilities: queries that are correctly responded at earlier training stages may become incorrectly responded as optimization proceeds, even the overall performance continues to improve. Further analysis reveals that this degradation coincides with a deterioration of the geometry of hidden representations, characterized by reduced intra-cluster compactness and diminished inter-cluster separation. Motivated by this observation, we propose Representation Geometry Regularization (RGR), which preserves representation geometry during post-training by encouraging compactness within cluster and separation across cluster. RGR can be seamlessly integrated into existing OPSD methods without altering their core optimization procedures. Extensive experiments on multiple OPSD methods and benchmarks show that RGR generally alleviates capability degradation while retaining capability improvements. Our code is available in https://github.com/supermachinelearner/RGR.", "url": "https://wpnews.pro/news/mitigating-capability-degradation-in-llm-self-improvement-via-representation", "canonical_source": "https://arxiv.org/abs/2610.12855", "published_at": "2026-10-12 04:00:00+00:00", "updated_at": "2026-10-12 04:41:10.948484+00:00", "lang": "en", "topics": ["large-language-models", "machine-learning", "ai-research", "artificial-intelligence"], "entities": ["arXiv", "Representation Geometry Regularization", "On-Policy Self-distillation", "supermachinelearner/RGR"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/mitigating-capability-degradation-in-llm-self-improvement-via-representation", "markdown": "https://wpnews.pro/news/mitigating-capability-degradation-in-llm-self-improvement-via-representation.md", "text": "https://wpnews.pro/news/mitigating-capability-degradation-in-llm-self-improvement-via-representation.txt", "jsonld": "https://wpnews.pro/news/mitigating-capability-degradation-in-llm-self-improvement-via-representation.jsonld"}}