Ziming Liu explains loss spikes: only the middle phase spikes as effective LR crosses critical LR
ZimingLiu11 · x · 2026-11-01
In a research thread, Ziming Liu explains why training loss spikes occur only in the middle phase: effective learning rate cycles — it rises past the critical learning rate, triggers a spike, drops, then recovers. Phase-boundary slopes near 1 suggest the loss-landscape bottom can't be approximated by a quadratic model (which would give slope 3).
More from Research
- Anthropic's FlashProve brings provable inference to LLMs with <100x overhead on Llama4 — AccBalanced · 2026-11-04
- Paperclip builds biomedical datasets in one shot, cutting months-long curation to about a minute — james_y_zou · 2026-11-04
- Embodied Agent Arena: Closed-Weight VLMs Lead All Five Robot Task Domains, Open-Weight Models Lag — Haojian Huang · 2026-11-04
- FPT University Paper Shows RAG Improves Stock Market Information Retrieval Accuracy and Latency — Huan Phong Thai · 2026-11-04
- Alibaba's adaptive GUI agent training cuts environment interactions by 56-61% — Yuxiang Chai · 2026-11-04
- Large-Scale Study: AI Reviewers Often Ignore How Much You Improved Your Paper — umd-zhou-lab · 2026-11-04