AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

Understanding SAM through Minimax Perspective

Ying Chen, Aoxi Li, Javad Lavaei

ICML 2026 · Optimization › Implicit bias and generalization · approach: Sharpness-aware minimization theory

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
add multiple inner ascent steps (M=5 to M=10) in the final phase at rho 0.2 (ablation)Periodic MSAM · CIFAR-100, ResNet-20test accuracy: 69.76% -> 70.32%Improved
use larger final-phase radius (0.2 instead of 0.1) with M=10 (ablation)Periodic MSAM · CIFAR-100, ResNet-20test accuracy: 70.01% -> 70.32%Improved
replace standard SAM with Periodic MSAM (M=10, rho 0.2 in last 20% epochs)SAM · CIFAR-100, ResNet-20test accuracy: 69.32% -> 70.32%Improved
replace SAM with Periodic MSAM (rho 0.2, M=5) in last 20% of trainingSAM rho 0.1 · Imagenettetop-1 validation accuracy: 75.01% -> 76.87%Improved
use multi-step MSAM (M=5) instead of single-step SAM at rho 0.10SAM · CIFAR-100, ResNet56test accuracy: 74.2 -> 74.8Improved
use multi-step MSAM (M=5) instead of SAM at rho 0.05SAM · CIFAR-100, ResNet56test accuracy: 75.2 -> 75.2No clear effect
replace SGD with SAM/MSAMSGD · CIFAR-100, ResNet56 rho 0.10test accuracy: 74.4 -> 74.2 (SAM), 74.8 (MSAM)Mixed
replace SGD with MSAM (M=3, rho 0.10)SGD · CIFAR-10, ResNet20test accuracy: 93.2 -> 93.7Improved
increase inner steps from M=5 to M=10 at small rho 0.025 (ablation)MSAM · CIFAR-10, ResNet20test accuracy: 93.6 -> 93.3Hurt
use MSAM (M=5) instead of standard SAM at rho 0.10SAM · CIFAR-10, WRN28-2test accuracy: 95.6 -> 95.8Improved
use MSAM (M=5) instead of SAM at rho 0.05SAM · CIFAR-10, ResNet56test accuracy: 95.3 -> 95.2No clear effect
replace single-step SAM with MSAM (M=5) at large rho 0.4SAM · Matrix completion n=100final loss: 1.5027 -> 7.5121e-05Improved

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