AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Cautious Weight Decay

Lizhang Chen, Jonathan Li, Kaizhao Liang, Baiyu Su, Cong Xie, Chen Liang, Ni Lao, qiang liu

ICLR 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging · 8 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
apply weight decay only where update and parameter signs align (CWD)AdamW · OLMo-1B 100B tokensvalidation loss: 2.65 -> 2.56Improved
apply CWDMuon · OLMo-1B 100B tokensvalidation loss: 2.51 -> 2.42Improved
replace sign mask with random Bernoulli mask of matched sparsity (ablation)AdamW+CWD · OLMo-1B 100B tokensvalidation loss: 2.56 -> 2.82Hurt
replace update-based mask with gradient-based mask (ablation)Muon+CWD · OLMo-1B 100B tokensvalidation loss: 2.42 -> 2.74Hurt
remove weight decay entirely (lambda=0) (ablation)AdamW · OLMo-1B 100B tokensvalidation loss: 2.65 -> 2.70Hurt
apply CWD to AdamW, Lion, Muon without retuningViT-S/16 · ImageNet validationaccuracy (%): 78.84 -> 79.45 (AdamW)Improved
apply CWD to MuonViT-B/16 · ImageNet validationaccuracy (%): 80.83 -> 81.04Improved
apply CWD to AdamW at 1B scaleOLMo-1B · downstream zero-shotHellaSwag acc_norm: 0.38 -> 0.40Improved
apply CWD to Muon at 1B scaleOLMo-1B · downstream zero-shotARC-Easy acc_norm: 0.51 -> 0.51No clear effect
use elementwise CWD instead of LoRA baselineTinyLlama · MMLU 5-shotaccuracy (%): 25.81 -> 26.42Improved
use elementwise CWD instead of SPDMistral-7B · WinoGrande 5-shotaccuracy (%): 78.81 -> 78.92No clear effect
use layerwise inner-product mask instead of elementwise mask (ablation)CWD on Mistral-7B · MMLU 5-shotaccuracy (%): 62.13 -> 61.76Hurt

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