← Back to all news
arXiv AI

Mitigating Capability Degradation in LLM Self-Improvement via Representation Geometry Regularization

The paper introduces Representation Geometry Regularization (RGR) to address capability degradation observed in On-Policy Self-distillation (OPSD) for large language models. OPSD can cause previously correct responses to become incorrect even as overall performance rises, a problem linked to deteriorating hidden representation geometry—specifically, less compact intra-cluster and weaker inter-cluster separation. RGR preserves this geometry by encouraging cluster compactness and separation during post‑training, and can be added to existing OPSD methods without changing their core optimization, with experiments showing it mitigates degradation while maintaining improvements.

Related stories

arXiv AI
Aug 19

Self-Distillation as a Performance Recovery Mechanism for LLMs: Counteracting Compression and Catastrophic Forgetting

The paper proposes Self‑Distillation Fine‑Tuning (SDFT) as a method to recover performance in Large Language Models that has been degraded by catastrophic forgetting, quantization, or pruning. It shows that SDFT restores model capabilities by aligning the high‑dimensional manifold of the student model’s hidden layers with that of a teacher model, as measured by Centered Kernel Alignment (CKA). The authors provide both empirical evidence of strong correlation between manifold alignment and performance recovery and a theoretical explanation linking generative capability to the structure of these manifolds.

By Chi Liu, Xin Chen, Xu Zhou, Fangbo Tu, Srinivasan Manoharan
llmsfine-tuningefficiencysafety
More like this →
arXiv AI
Oct 1

Mitigating Memorization In Language Models

The paper explores ways to reduce the memorization of training data in language models, testing three regularizer-based, three finetuning-based, and eleven machine unlearning methods—five of which are newly introduced. It introduces TinyMem, a lightweight suite of small models for rapid testing of these mitigation techniques, and shows that unlearning methods, particularly BalancedSubnet, outperform others in removing memorized content while maintaining task performance. The study also finds that regularizer-based approaches are slow and ineffective, while finetuning methods are costly, especially when high accuracy is required.

By Mansi Sakarvadia, Aswathy Ajith, Arham Khan, Nathaniel Hudson, Caleb Geniesse, Kyle Chard, Yaoqing Yang, Ian Foster, Michael W. Mahoney
llmsfine-tuning
More like this →
arXiv Machine Learning
Oct 15

How Inner Optimizers Shape Replica Scaling

The paper investigates how different inner optimizers affect replica scaling when training large language models across distributed compute resources. It identifies two key factors—covariance of pseudo-gradients and batch-size impact on noise scale—that influence performance as replicas increase. The authors show that whitening-based pre‑conditioning as an inner optimizer yields strong replica scaling, improving performance relative to DP as model size and data grow.

By Eugene Belilovsky, Stefan Horoi, Edouard Oyallon, Benjamin Th\'erien
llms
More like this →
arXiv AI
Jun 2

Memory-Efficient LLM Training with Dynamic Sparsity: From Stability to Practical Scaling

arXiv:2606. 00888v1 Announce Type: cross Abstract: Dynamic Sparse Training (DST) offers a promising paradigm for improving the training and inference efficiency of deep neural networks; however, we find that in large language model training, DST can suffer from optimization instability, manifested as loss spikes after topology updates.

By Qiao Xiao, Boqian Wu, Patrik Okanovic, Tomasz Sternal, Maurice van Keulen, Elena Mocanu, Mykola Pechenizkiy, Decebal Constantin Mocanu, Torsten Hoefler
llmsefficiency
More like this →