AI Paper Atlas

Research map / Optimization / Optimizers

Per-example Gradients: a New Frontier for Understanding and Improving Optimizers

Vincent Roulet, Atish Agarwala

ICML 2026 · Optimization › Optimizers · approach: Memory-efficient LLM optimizers · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
place sign last (SignEMA) instead of before averaging (MicroSignSGD) (ablation)signSGD variants · C4 (151M decoder LM, B=64)training lossImproved
place sign after averaging but before EMA (SignSGD) instead of per-example sign (ablation)signSGD variants · C4 (151M decoder LM, B=64)training lossImproved
use per-example mean of squared gradients as preconditioner (MicroAdam)Adam · C4 (151M decoder LM, B=64)eval loss: 3.06 vs 3.10Hurt
use variance-only preconditioner (MicroAdamVar) (ablation)MicroAdam · C4 (151M decoder LM, B=64)training lossHurt
use ReLU-filtered mean-squared preconditioner with clipping (MicroAdamMSQ)Adam · C4 (151M decoder LM, B=64)eval loss: 3.06 -> 3.05Improved
add gradient clipping to MicroAdam (ablation)MicroAdam · C4 (151M decoder LM, B=64)training spikesImproved
use unfiltered mean-squared estimator without ReLU (ablation)MicroAdamMSQ · C4 (151M decoder LM, B=64)stabilityHurt
use per-example mean-squared preconditioner (MicroAdamMSQ) for computeAdam · C4 (151M decoder LM, B=64)wall-clock time: 20% more wall-clock timeHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy