New method helps large AI models remember while learning new tasks
Entropic Control for Homeostatic Optimization: Pre-empting Catastrophic Forgetting via Dynamic Feedback
Machine LearningArtificial IntelligenceComputation and Language
Summary
When large AI language models are fine-tuned with new data, they often forget what they previously learned, which is called catastrophic forgetting. The authors propose a technique called ECHO that watches a measure called entropy to detect when the model might forget important information. By adjusting training based on this signal, ECHO helps the model keep a good balance between learning new things and retaining old knowledge. Tests on models of different sizes show that ECHO improves this balance and works well across various training setups.
What this means in practice
- •For ml engineers: Maintain pretrained model knowledge better during supervised fine-tuning to improve reliability and reduce re-training costs.
- •For ai infrastructure teams: Adjust training workflows dynamically to prevent forgetting risks without expensive reference model computations.
Authors
Shangxi Wu, Xiaorong Yan, Xiaojun Meng, Yuxiang Zhang, Jiansheng Wei
Abstract
Supervised Fine-Tuning (SFT) of LLMs suffers from catastrophic forgetting, and prevailing mitigation strategies struggle to reconcile the plasticity-stability dilemma. To address this, we propose Entropic Control for Homeostatic Optimization (ECHO), a closed-loop control framework. By decomposing the KL divergence, we identify output entropy as a dominant observable component of KL variation. This enables us to leverage entropy as a computationally efficient online proxy to monitor forgetting risks in real-time, bypassing expensive reference-model inference. Consequently, ECHO employs an elastic controller that dynamically modulates the gradient weight of replay data via a proportional control law, using deviations from the pre-trained entropic equilibrium as an error signal. Empirical evaluations across 0.6B to 9B parameter architectures demonstrate that ECHO lies on, and in several comparisons extends, the observed plasticity-stability frontier and maintains high operational robustness against varying data mixing ratios.