AI News

The paper investigates capability degradation in on-policy self-distillation (OPSD), a self-improvement setting in which the same large language model acts as teacher and student under different conditioning contexts. The authors find that optimization can make previously correct queries become incorrect even while aggregate performance continues to rise. This degradation is associated with deterioration in hidden-representation geometry, including lower intra-cluster compactness and weaker inter-cluster separation. They propose Representation Geometry Regularization (RGR), which encourages representations from the same cluster to remain compact while separating different clusters during post-training. RGR can be added to existing OPSD methods without changing their core optimization procedures. Experiments across multiple OPSD methods and benchmarks show that it generally reduces capability degradation while preserving capability gains. The authors also release implementation code.