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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| 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 7B | average accuracy: 44.5 -> 61.5 | Improved |
| fine-tune DOT-MoE on 1.2B tokens after alignment | DOT-MoE · common-sense average, LLaMA-2 7B | average accuracy: 61.5 -> 66.6 | Improved |
| scale fine-tuning from 1.2B to 7B tokens | DOT-MoE · common-sense average, LLaMA-3 8B | average accuracy: 67.8 -> 71.0 | Improved |
| align at 75% FFN sparsity rather than 50% (ablation) | DOT-MoE · Qwen2.5-7B avg accuracy at 75% inference sparsity | average accuracy: 45.13 -> 55.64 | Improved |
| align at 75% FFN sparsity rather than 50% (ablation) | DOT-MoE · Qwen2.5-7B avg accuracy at 30% inference sparsity | average accuracy: 71.21 -> 71.78 | No clear effect |
| increase number of experts (granularity) (ablation) | DOT-MoE · Qwen2.5 7B WikiText perplexity at 25% active | perplexity | Mixed |
| increase expert count from 8 to 148 at fixed active parameters (ablation) | DOT-MoE · inference throughput with vLLM fused MoE | throughput | No clear effect |
| use E=8 top-k=2 instead of E=148 (ablation) | DOT-MoE · Qwen2.5-7B average accuracy | average accuracy: 67.2 vs. 67.7 | No clear effect |
| CMoE with 37 experts and 9 active on Qwen2.5-7B (ablation) | CMoE · WikiText perplexity | perplexity: >5K PPL | Hurt |
| OT-based head assignment instead of random assignment for attention experts (ablation) | attention MoEfication · Qwen2.5-7B, 50% attention sparsity | average accuracy: 46.2 -> 64.1 | Improved |
| 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 similarity | cosine similarity: 0.55 -> 0.84 | Improved |
| use DOT-MoE assignment instead of CMoE on a 32B model | Qwen2.5-32B · common-sense average, 25% active | average accuracy: 38.8 -> 73.1 | Improved |
Closest papers by meaning
- Advancing Expert Specialization for Better MoE (NeurIPS 2025 · similarity 0.91)
- Pruning and Distilling Mixture-of-Experts into Dense Language Models (NeurIPS 2026 · similarity 0.90)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.90)
- BAM! Just Like That: Simple and Efficient Parameter Upcycling for Mixture of Experts (NeurIPS 2024 · similarity 0.90)
- Layerwise Recurrent Router for Mixture-of-Experts (ICLR 2025 · similarity 0.90)
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · similarity 0.89)
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.
