# Mechanism-Driven Monitors for Preemptive Detection of LLM Training Instability

arXiv: https://arxiv.org/abs/2606.28116 · HTML: https://mlsanity.com/paper/2606.28116/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Builds on 25 of its 33 references that are in the map.
- First posted on arXiv 26 Jun 2026.

## Builds on (most cited first, 25 of 25)

- [Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer](https://mlsanity.com/paper/1701.06538/) (2017)
- [Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity](https://mlsanity.com/paper/2101.03961/) (2021)
- [Neural Tangent Kernel: Convergence and Generalization in Neural Networks](https://mlsanity.com/paper/1806.07572/) (2018)
- [GLM-130B: An Open Bilingual Pre-trained Model](https://mlsanity.com/paper/2210.02414/) (2022)
- [Wide Neural Networks of Any Depth Evolve as Linear Models Under Gradient Descent](https://mlsanity.com/paper/1902.06720/) (2019)
- [Scaling Vision Transformers to 22 Billion Parameters](https://mlsanity.com/paper/2302.05442/) (2023)
- [On Lazy Training in Differentiable Programming](https://mlsanity.com/paper/1812.07956/) (2018)
- [GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection](https://mlsanity.com/paper/2403.03507/) (2024)
- [Gradient Descent on Neural Networks Typically Occurs at the Edge of Stability](https://mlsanity.com/paper/2103.00065/) (2021)
- [Attention is Not All You Need: Pure Attention Loses Rank Doubly Exponentially with Depth](https://mlsanity.com/paper/2103.03404/) (2021)
- [PyHessian: Neural Networks Through the Lens of the Hessian](https://mlsanity.com/paper/1912.07145/) (2019)
- [Auxiliary-Loss-Free Load Balancing Strategy for Mixture-of-Experts](https://mlsanity.com/paper/2408.15664/) (2024)
- [Small-scale proxies for large-scale Transformer training instabilities](https://mlsanity.com/paper/2309.14322/) (2023)
- [Stabilizing Transformer Training by Preventing Attention Entropy Collapse](https://mlsanity.com/paper/2303.06296/) (2023)
- [Low-Rank Bottleneck in Multi-head Attention Models](https://mlsanity.com/paper/2002.07028/) (2020)
- [mHC: Manifold-Constrained Hyper-Connections](https://mlsanity.com/paper/2512.24880/) (2025)
- [Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models](https://mlsanity.com/paper/2601.07372/) (2026)
- [Attention Residuals](https://mlsanity.com/paper/2603.15031/) (2026)
- [Spike No More: Stabilizing the Pre-training of Large Language Models](https://mlsanity.com/paper/2312.16903/) (2023)
- [MaskLLM: Learnable Semi-Structured Sparsity for Large Language Models](https://mlsanity.com/paper/2409.17481/) (2024)
- [Approaching Deep Learning through the Spectral Dynamics of Weights](https://mlsanity.com/paper/2408.11804/) (2024)
- [Self-attention Networks Localize When QK-eigenspectrum Concentrates](https://mlsanity.com/paper/2402.02098/) (2024)
- [Dissecting Query-Key Interaction in Vision Transformers](https://mlsanity.com/paper/2405.14880/) (2024)
- [Why Low-Precision Transformer Training Fails: An Analysis on Flash Attention](https://mlsanity.com/paper/2510.04212/) (2025)
- [A Scalable Measure of Loss Landscape Curvature for Analyzing the Training Dynamics of LLMs](https://mlsanity.com/paper/2601.16979/) (2026)

## Built on by (most cited first, 2 of 2)

- [One QK Channel, Many Sources: Tracing Low-Precision Attention Collapse](https://mlsanity.com/paper/2608.02091/) (2026)
- [Delta-Matching: Closing the Final Gap of Native 8-bit Training for LLMs](https://mlsanity.com/paper/2609.37852/) (2026)

## Abstract (arXiv)

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.