# Adaptive Preconditioners Trigger Loss Spikes in Adam

arXiv: https://arxiv.org/abs/2506.04805 · HTML: https://mlsanity.com/paper/2506.04805/

## Key facts

- Cited by 13 papers in the map (data through 14 Oct 2026).
- Builds on 40 of its 49 references that are in the map.
- First posted on arXiv 5 Jun 2025.

## Builds on (most cited first, 30 of 40)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [Llama 2: Open Foundation and Fine-Tuned Chat Models](https://mlsanity.com/paper/2307.09288/) (2023)
- [PaLM: Scaling Language Modeling with Pathways](https://mlsanity.com/paper/2204.02311/) (2022)
- [On the Convergence of Adam and Beyond](https://mlsanity.com/paper/1904.09237/) (2019)
- [On the Variance of the Adaptive Learning Rate and Beyond](https://mlsanity.com/paper/1908.03265/) (2019)
- [CogView: Mastering Text-to-Image Generation via Transformers](https://mlsanity.com/paper/2105.13290/) (2021)
- [Gradient Descent on Neural Networks Typically Occurs at the Edge of Stability](https://mlsanity.com/paper/2103.00065/) (2021)
- [The large learning rate phase of deep learning: the catapult mechanism](https://mlsanity.com/paper/2003.02218/) (2020)
- [On the Convergence of A Class of Adam-Type Algorithms for Non-Convex Optimization](https://mlsanity.com/paper/1808.02941/) (2018)
- [A Simple Convergence Proof of Adam and Adagrad](https://mlsanity.com/paper/2003.02395/) (2020)
- [Stabilizing Transformer Training by Preventing Attention Entropy Collapse](https://mlsanity.com/paper/2303.06296/) (2023)
- [On the Convergence of Stochastic Gradient Descent with Adaptive Stepsizes](https://mlsanity.com/paper/1805.08114/) (2018)
- [Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability](https://mlsanity.com/paper/2209.15594/) (2022)
- [The Break-Even Point on Optimization Trajectories of Deep Neural Networks](https://mlsanity.com/paper/2002.09572/) (2020)
- [Understanding Gradient Descent on Edge of Stability in Deep Learning](https://mlsanity.com/paper/2205.09745/) (2022)
- [A Sufficient Condition for Convergences of Adam and RMSProp](https://mlsanity.com/paper/1811.09358/) (2018)
- [Adam Can Converge Without Any Modification On Update Rules](https://mlsanity.com/paper/2208.09632/) (2022)
- [Adaptive Gradient Methods at the Edge of Stability](https://mlsanity.com/paper/2207.14484/) (2022)
- [On the Convergence of Adaptive Gradient Methods for Nonconvex Optimization](https://mlsanity.com/paper/1808.05671/) (2018)
- [On the Relation Between the Sharpest Directions of DNN Loss and the SGD Step Length](https://mlsanity.com/paper/1807.05031/) (2018)
- [Understanding the unstable convergence of gradient descent](https://mlsanity.com/paper/2204.01050/) (2022)
- [Understanding the Generalization Benefit of Normalization Layers: Sharpness Reduction](https://mlsanity.com/paper/2206.07085/) (2022)
- [A Walk with SGD](https://mlsanity.com/paper/1802.08770/) (2018)
- [Understanding Optimization in Deep Learning with Central Flows](https://mlsanity.com/paper/2410.24206/) (2024)
- [Analyzing Sharpness along GD Trajectory: Progressive Sharpening and Edge of Stability](https://mlsanity.com/paper/2207.12678/) (2022)
- [A Theory on Adam Instability in Large-Scale Machine Learning](https://mlsanity.com/paper/2304.09871/) (2023)
- [Beyond the Quadratic Approximation: the Multiscale Structure of Neural Network Loss Landscapes](https://mlsanity.com/paper/2204.11326/) (2022)
- [In Search of Adam's Secret Sauce](https://mlsanity.com/paper/2505.21829/) (2025)
- [A Qualitative Study of the Dynamic Behavior for Adaptive Gradient Algorithms](https://mlsanity.com/paper/2009.06125/) (2020)
- [Linear Convergence of Adaptive Stochastic Gradient Descent](https://mlsanity.com/paper/1908.10525/) (2019)

## Built on by (most cited first, 13 of 13)

- [Towards Understanding Adam Convergence on Highly Degenerate Polynomials](https://mlsanity.com/paper/2603.09581/) (2026)
- [Grokking or Glitching? How Low-Precision Drives Slingshot Loss Spikes](https://mlsanity.com/paper/2605.06152/) (2026)
- [Adapt or Forget: Provable Tradeoffs Between Adam and SGD in Nonstationary Optimization](https://mlsanity.com/paper/2605.04269/) (2026)
- [Weight-norm Criticality: A Mechanism for Loss Spikes Induced by the Normalization and Weight Decay](https://mlsanity.com/paper/2607.21005/) (2026)
- [Finite-Horizon Input-Output Dynamics of Minibatch Perturbations in AdamW](https://mlsanity.com/paper/2608.19762/) (2026)
- [Adam at the Edge of Stability: Adaptive Feedback, Provable Oscillation, and Gradient Reversal](https://mlsanity.com/paper/2608.20638/) (2026)
- [Self-Supervised Learning with a Multi-Task Latent Space Objective](https://mlsanity.com/paper/2602.05845/) (2026)
- [One Algorithm, Two Goals: Dual Scoring for Parameter and Data Selection in LLM Fine-Tuning](https://mlsanity.com/paper/2605.06166/) (2026)
- [SCAPE: Accurate and Efficient LLM Training with Extreme Sparse Communication](https://mlsanity.com/paper/2607.01678/) (2026)
- [On the Principles Behind Neural Network Optimizers](https://mlsanity.com/paper/2608.16760/) (2026)
- [Beyond Quadratic Loss: The Stability Phase Diagram of Adam](https://mlsanity.com/paper/2609.18314/) (2026)
- [Simulating Training Spikes of Deep Neural Networks by Layer-wise Random Feature Models](https://mlsanity.com/paper/2610.13524/) (2026)
- [Minimax Optimality of Shampoo for Nonstationary Stochastic Optimization](https://mlsanity.com/paper/2610.16923/) (2026)

## Abstract (arXiv)

Loss spikes commonly emerge during neural network training with the Adam optimizer across diverse architectures and scales, yet their underlying mechanism remains elusive. While previous explanations attribute these phenomena to sharper loss landscapes at lower loss, we show that landscape geometry alone is insufficient to explain the phenomenon. In this work, we pinpoint the root cause in the internal dynamics of Adam's second moment estimator. We identify a critical ``decoupling'' mechanism where the adaptive preconditioner vt fails to track the instantaneous squared gradients gt2, causing the adaptive mechanism to effectively fail. This decoupling allows the preconditioner to decay autonomously despite rising gradients, which pushes the maximum eigenvalue of the preconditioned Hessian beyond the stability threshold 2/η for sustained periods, manifesting as dramatic loss spikes. Through a quadratic approximation analysis, we theoretically and experimentally characterize five distinct stages of spike evolution and propose a predictor for anticipating spikes based on gradient-directional curvature. We empirically find that the proposed loss spike mechanism, although derived from simplified models, generalizes well to practical scenarios ranging from small neural networks to large-scale Transformers.