AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

Libra: Effective yet Efficient Load Balancing for Large-scale MoE Inference

Jaehoon Yang, Yushin Kim, Seokwon Moon, Yeonhong Park, Jae W. Lee

ICLR 2026 · Architectures and efficiency › Mixture of experts · approach: Expert routing and specialization

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
replace Lina's lookup-table predictor with lookahead gating predictionLibra · BookCorpus (Qwen3MoE)expert prediction accuracy (%): 47.3 -> 91.7Improved
replace Lina's lookup-table predictor with lookahead gating predictionLibra · BookCorpus (GLM-4.5)expert prediction accuracy (%): 11.7 -> 79.6Improved
replace Lina's lookup-table predictor with lookahead gating predictionLibra · HellaSwag (GLM-4.5)expert prediction accuracy (%): 11.5 -> 76.6Improved
use Libra's predictor instead of Lina'sMoE layer · MoE layer breakdownpredict overhead: 0.40 -> 0.03Improved
add Libra load balancing and overlapped execution over SGLangQwen3MoE MoE layer · MoE layer breakdown (seq 1024, batch 32)total MoE layer latency: 13.61 -> 9.07Improved
use Libra instead of LinaQwen3MoE MoE layer · MoE layer breakdown (seq 1024, batch 32)total MoE layer latency: 11.33 -> 9.07Improved
add Libra over SGLang at long context and large batchQwen3MoE · End-to-end TTFTTTFT reduction (%): up to 27%Improved
add Libra over SGLang at short sequences and small batchQwen3MoE · End-to-end latency, batch 8, input 1024, output 1latency (s): 0.36 -> 0.38Hurt
add Libra over SGLang at batch 64, input 1024Qwen3MoE · End-to-end latency, output 1latency (s): 2.72 -> 2.00Improved
add Libra over Lina at batch 64, input 2048Qwen3MoE · End-to-end latency, output 1latency (s): 4.76 -> 4.00Improved
add Libra to improve prefill throughput over state of the artQwen3MoE and GLM-4.5 · Prefill throughput, 4 imbalanced datasetsthroughput (tokens/s): up to 19.2%Improved

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