AI Paper Atlas

Research map / Architectures and efficiency / Mixture of experts

DOT-MoE: Differentiable Optimal Transport for MoEfication

Udbhav Bamba, Arnav Chavan, Aryamaan Thakur, Steven Teig, Deepak Gupta

ICML 2026 · Architectures and efficiency › Mixture of experts · approach: Efficient MoE language model systems · 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
learn neuron-to-expert assignment by optimal transport instead of CMoE co-activation clustering (zero-shot)dense-to-MoE conversion · common-sense average, LLaMA-2 7Baverage accuracy: 44.5 -> 61.5Improved
fine-tune DOT-MoE on 1.2B tokens after alignmentDOT-MoE · common-sense average, LLaMA-2 7Baverage accuracy: 61.5 -> 66.6Improved
scale fine-tuning from 1.2B to 7B tokensDOT-MoE · common-sense average, LLaMA-3 8Baverage accuracy: 67.8 -> 71.0Improved
align at 75% FFN sparsity rather than 50% (ablation)DOT-MoE · Qwen2.5-7B avg accuracy at 75% inference sparsityaverage accuracy: 45.13 -> 55.64Improved
align at 75% FFN sparsity rather than 50% (ablation)DOT-MoE · Qwen2.5-7B avg accuracy at 30% inference sparsityaverage accuracy: 71.21 -> 71.78No clear effect
increase number of experts (granularity) (ablation)DOT-MoE · Qwen2.5 7B WikiText perplexity at 25% activeperplexityMixed
increase expert count from 8 to 148 at fixed active parameters (ablation)DOT-MoE · inference throughput with vLLM fused MoEthroughputNo clear effect
use E=8 top-k=2 instead of E=148 (ablation)DOT-MoE · Qwen2.5-7B average accuracyaverage accuracy: 67.2 vs. 67.7No clear effect
CMoE with 37 experts and 9 active on Qwen2.5-7B (ablation)CMoE · WikiText perplexityperplexity: >5K PPLHurt
OT-based head assignment instead of random assignment for attention experts (ablation)attention MoEfication · Qwen2.5-7B, 50% attention sparsityaverage accuracy: 46.2 -> 64.1Improved
OT assignment instead of CMoE under identical fine-tuning, representation similarity (ablation)dense-to-MoE conversion · Qwen2.5-7B (E=8, top-k=2), pre-LM cosine similaritycosine similarity: 0.55 -> 0.84Improved
use DOT-MoE assignment instead of CMoE on a 32B modelQwen2.5-32B · common-sense average, 25% activeaverage accuracy: 38.8 -> 73.1Improved

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