AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

DAG-MoE: From Simple Mixture to Structural Aggregation in Mixture-of-Experts

Jiarui Feng, Hanqing Zeng, Karish Grover, Ruizhong Qiu, Yinglong Xia, Qiang Zhang, Qifan Wang, Ren Chen, Dongqi Fu, Jiayi Liu, Zhuokai Zhao, Xiangjun Fan and 2 more

ICML 2026 · Architectures and efficiency › Mixture of experts · approach: Expert routing and specialization · 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
add DAG learning module (structural aggregation) to MoEMoE-l · Pile / Wiki / FineWeb-Edu / C4 (40B tokens)perplexity: 10.51 -> 10.27 (Pile); 35.21 -> 34.21 (C4)Improved
add DAG module instead of Chain-of-Experts iterationCoE · Pile 12B tokens, matched paramsperplexity reduction: 0.480 -> 0.587Improved
add DAG module instead of a shared expertStandard MoE · Pile 12B tokens, matched paramsperplexity reduction: 0.433 -> 0.587Improved
add structured DAG module in place of naive MLP mixing (d_g=32) (ablation)DAG-MoE-s · Eval loss, 12B tokenseval loss: 2.8406 -> 2.6948Improved
add naive MLP mixing of expert outputs to standard MoE (d_g=64) (ablation)Standard MoE · Eval loss, 12B tokenseval loss: 2.7168 -> 2.8006Hurt
increase DAG iterations from L=1 to L=2 (d_g=32) (ablation)DAG-MoE-s · Eval loss, 12B tokenseval loss: 2.6948 -> 2.6899Improved
increase DAG iterations L from 2 to 3 (ablation)DAG-MoE-s/m · Pile subset, 12B tokensperplexityNo clear effect
add a single DAG iteration (d_g=64, L=1) (ablation)DAG-MoE-s · Pile subset, top-K 4 and 8perplexity: about 0.5 reduction in perplexityImproved
use d_g=64,L=2 rather than d_g=128,L=1 (ablation)DAG-MoE-s · Pile subsetperplexity: d_g=64,L=2 outperforms d_g=128,L=1Improved
add DAG module (L=1) vs standard MoEDAG-MoE-s · Training throughputtokens/s: 0.7744 -> 0.7627Hurt
add DAG module (L=2) vs standard MoEDAG-MoE-s · Training throughputtokens/s: 0.7744 -> 0.7404Hurt
add DAG module after instruction fine-tuningMoE-l · GPQAaccuracy: 21.72 -> 27.78Improved

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