AI Paper Atlas

Research map / Optimization / Distributed and compressed

DES-LOC: Desynced Low Communication Adaptive Optimizers for Foundation Models

Alex Iacob, Lorenzo Sani, Mher Safaryan, Paris Giampouras, Samuel Horváth, Andrej Jovanovic, Meghdad Kurmanji, Preslav Aleksandrov, William F. Shen, Xinchi Qiu, Nicholas D. Lane

ICLR 2026 · Optimization › Distributed and compressed · approach: Gradient quantization and sparsification

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
synchronize parameters less frequently (larger Kx) (ablation)DES-LOC (ADOPT, 135M) · SmolLM2 mixture, 135M modelperplexityHurt
synchronize second momentum less frequently (larger Kv) (ablation)DES-LOC (ADOPT, beta2=0.9999) · SmolLM2 mixture, 135M modelperplexityNo clear effect
synchronize first momentum more frequently (ablation)DES-LOC (ADOPT, beta1=0.95) · SmolLM2 mixture, 135M modelperplexityMixed
desync momenta: Ku=3Kx and Kv=6Kx instead of Local Adam syncing all states at KDES-LOC · SmolLM2 mixture, 135M modelcommunication cost at matched perplexity: 2x lower communication than Local AdamImproved
sync first momentum less often than second (Ku=6Kx, Kv=3Kx) (ablation)DES-LOC-ADOPT · 135M, 2x lower communication than Local AdamperplexityHurt
set Ku=Kv=4Kx (ablation)DES-LOC-ADOPT · 135M, 2x lower communication than Local AdamperplexityMixed
split communication between momenta (Ku vs Kv) (ablation)DES-LOC-Adam (beta1=beta2=0.95) · 135M, 2x lower communicationperplexityNo clear effect
sync optimizer states eventually instead of keeping them local or resetting themDES-LOC vs FAVG+OPT · 1.7B, 40B tokensICL average: 50.7 -> 51.6Improved
keep optimizer states local (FedAvg + persistent states) (ablation)FedAvg · 1.7B long horizontraining stabilityHurt
double worker count at step 1536DES-LOC and Local Adam vs heuristic baselines · 135M modelperplexity and gradient normsNo clear effect
use DES-LOC or Local Adam instead of DDP at billion scale1.7B model · ICL averageICL average: 53.8 -> 51.6Hurt

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