Research map / Optimization / Optimizers
Adaptive Preconditioners Trigger Loss Spikes in Adam
Zhiwei Bai, Zhangchen Zhou, Jiajie Zhao, Xiaolong Li, Zhiyu li, Feiyu Xiong, Hongkang Yang, Yaoyu Zhang, Zhi-Qin John Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase epsilon to 0.1 at spike onset (ablation) | Adam · FNN function approximation | loss spike | Improved |
| reduce beta2 from 0.999 to 0.9 (ablation) | Adam · Transformer 10M synthetic reasoning | loss spike frequency | Improved |
| reduce beta2 (0.999, 0.9, 0.85) (ablation) | Adam · LLaMA 187M on SlimPajama | loss spike frequency: 0.999 -> 0.9 -> 0.85 | Improved |
| use low beta1=0.6 and beta2=0.5 (tight coupling of v_t and gradient) (ablation) | Adam · 1D quadratic | loss behavior: spike -> EoS oscillation | Improved |
| increase beta1 to reduce preconditioned curvature (ablation) | Adam · 1D quadratic | loss spikes | No clear effect |
| use gradient-directional curvature lambda_grad instead of lambda_max as spike predictor (ablation) | Adam · two-layer net sin(x)+sin(4x) | spike prediction: lambda_max exceeds 2/eta at 10 steps vs 7 spikes | Improved |
| use sustained (min over 3 steps) lambda_grad predictor (ablation) | Adam · Transformer 10M | spike prediction: matches all 7 spikes | Improved |
| use learning rate decay schedules (ablation) | Adam · 1D quadratic (beta2 near 1) | spike avoidance | No clear effect |
| use Adam instead of gradient descent at small learning rate | Adam · 1D quadratic | loss stability | Hurt |
Closest papers by meaning
- The Expected Loss of Preconditioned Langevin Dynamics Reveals the Hessian Rank (AAAI 2024 · similarity 0.86)
- Exact risk curves of signSGD in High-Dimensions: quantifying preconditioning and noise-compression effects (ICML 2025 · similarity 0.86)
- The Hidden Ratio in Adam: Stable Structure, Compression, and Sign Dynamics (NeurIPS 2026 · similarity 0.86)
- Adam under Generalized Smoothness with Second-Moment-Type Stochastic Gradients (NeurIPS 2026 · similarity 0.86)
- Adam Reduces a Unique Form of Sharpness: Theoretical Insights Near the Minimizer Manifold (NeurIPS 2025 · similarity 0.85)
- Preconditioners for the Stochastic Training of Neural Fields (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
