Mechanism-Driven Monitors for Preemptive Detection of LLM Training Instability

2026-06-26

AI Summary

This work addresses the challenge of training failures in large language models, which often persist for thousands of optimization steps before manifesting as obvious loss divergence, leading to substantial computational waste. The authors propose a mechanism-aware, proactive monitoring approach that deploys internal detectors at the earliest points where failure signatures become measurable. Specifically, they introduce diagnostic signals grounded in the functional principles of critical modules—such as spectral entropy derived from the bilinear decomposition of QK matrices under low-precision Flash Attention and behavioral metrics of MoE router expert selection. By leveraging these module-specific indicators, the method enables early and accurate identification of diverse failure modes thousands of steps before loss divergence occurs, significantly outperforming conventional detection strategies based solely on loss values or gradient norms, particularly in scenarios involving low-precision attention, excessively high learning rates, or compound faults.

Abstract

Frontier large language model training consumes massive accelerator fleets and long wall-clock computation, making stability failures costly when they occur. After a numerical or a hyperparameter fault has already destabilized the training dynamics, it may continue for thousands of steps while loss and gradient norms still appear normal. We study mechanism-driven detection of training instability by deriving internal monitors from the functional role of each critical module and from the earliest computational sites where failures are expected to produce measurable signatures. For low-precision flash attention, we monitor the spectral entropy of a QK bilinear decomposition, whose first-order term becomes abnormal before the loss fully collapses. For MoE routers, we derive indicators from their role in expert selection. Our fault-injection experiments on low-precision attention, large learning-rate, and combined faults show that these signals provide distinct signatures for different failures, triggering thousands of steps before loss divergence.

Problem

• training instability
• large language models
• fault detection
• numerical errors
• hyperparameter faults

Innovation

• mechanism-driven monitoring
• training instability detection
• low-precision attention
• MoE routers
• spectral entropy