Research map / Architectures and efficiency / Mixture of experts
Monet: Mixture of Monosemantic Experts for Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Monet, an MoE pretraining architecture with 262,144 experts per layer via product-key composition of horizontal/vertical expert decompositions, enabling monosemantic experts and knowledge/toxicity unlearning by expert removal.
Problem
Polysemanticity in LLM neurons hinders mechanistic interpretability; post-hoc SAEs degrade LLM performance, and existing MoE designs have few experts or parameter counts that grow linearly with expert count.
Approach
Decompose each expert into bottom and top linear layers and compose N=K^2 experts from sqrt(N) bottom and sqrt(N) top weight sets (product key composition, with horizontal or vertical splits), so total parameters scale with sqrt(N) and sparse dictionary learning is done end-to-end during pretraining with uniformity and ambiguity losses.
Key results
- Monet-VD 1.4B reaches 0.510 average 5-shot accuracy on open-ended benchmarks vs 0.518 for the dense LLaMA 1.3B, while Gemma Scope SAE-augmented Gemma 2 2B drops to 0.473 (MLP SAE) and 0.337 (residual SAE) 0-shot.
- Purging up to 4.1% of experts on RealToxicityPrompts reduces toxic-prompt toxicity probability from 0.926 to 0.768 with average helpfulness unchanged (0.478).
- Deleting target-language experts in CodeMonet 1.4B drops pass@100 by up to 30.6 points for Python while other languages change by only about -0.6 to -1.8 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use vertical rather than horizontal expert decomposition (ablation) | Monet 1.4B · Open-ended benchmarks, 0-shot and 5-shot | Avg: 0.463 -> 0.478; 0.487 -> 0.510 | Helped |
| Mask toxic experts at threshold 0.1 (masking ratio 4.1%) | Monet · RealToxicityPrompts (toxic prompts) | Expected maximum toxicity: 0.795 -> 0.657 | Helped |
| Mask toxic experts at threshold 0.1 (masking ratio 4.1%) | Monet · RealToxicityPrompts (toxic prompts) | Toxicity probability: 0.926 -> 0.768 | Helped |
| Mask toxic experts at threshold 0.1 (masking ratio 4.1%) | Monet · RealToxicityPrompts | Avg. Performance (Helpfulness): 0.478 -> 0.478 | No clear effect |
| Mask toxic experts at threshold 0.1 (masking ratio 5.4%) | Monet · ToxiGen (Hate) | RoBERTa Score: 0.642 -> 0.504 | Helped |
| Mask toxic experts at threshold 0.1 (masking ratio 5.4%) | Monet · ToxiGen | Avg. Performance (Helpfulness): 0.478 -> 0.473 | Hurt |
| Set auxiliary loss weight λ to 1 × 10^-3 instead of omitting it (ablation) | Monet 1.4B · Open-ended benchmarks (5-shot) | Avg: 0.505 -> 0.510 | Helped |
| Set auxiliary loss weight λ to 5 × 10^-3 instead of omitting it (ablation) | Monet 1.4B · Open-ended benchmarks (5-shot) | Avg: 0.505 -> 0.502 | No clear effect |
| Reuse expert routing every 4 layers (ablation) | Monet 1.4B · Open-ended benchmarks (5-shot) | Avg: 0.518 -> 0.510 | Hurt |
| Use routing group size 1 (ablation) | Monet 1.4B · Open-ended benchmarks (5-shot) | Avg: 0.518 -> 0.511 | Hurt |
Surprising findings
- Gemma Scope SAEs substantially degrade a pretrained LLM's open-ended benchmark performance, especially the residual-stream SAE (0.337 average vs 0.571 for the base model).
- OLMoE showed no experts specialized by the skewness criterion in MMLU domains, with routing scores spread evenly.
Limitations
- Expert selection criteria (routing-score skewness and Pearson correlation with toxicity) are basic and minimal.
- Self-explained expert interpretations are only shown qualitatively, with no quantitative automated interpretability evaluation.
- Parametric knowledge manipulation is limited to unlearning by expert removal.
Experimental setup
Datasets: FineWeb-Edu StarCoderData C4 MMLU MMLU-Pro MULTIPL-E RealToxicityPrompts ToxiGen
Benchmarks: MMLU ARC WinoGrande PIQA SIQA OpenBookQA HellaSwag CommonsenseQA MULTIPL-E RealToxicityPrompts ToxiGen MMLU-Pro
Metrics: accuracy pass@100 expected maximum toxicity toxicity probability RoBERTa hate score
Baselines: LLaMA dense models OLMoE 6.9B Gemma 2 2B with Gemma Scope SAEs PEER
Builds on: Parameter Efficient Expert Retrieval (PEER) product key memory Sparse Autoencoders BatchTopK OLMoE
Closest papers by meaning
- Mixture of Experts Made Intrinsically Interpretable (ICML 2025 · similarity 0.88)
- MoEUT: Mixture-of-Experts Universal Transformers (NeurIPS 2024 · similarity 0.87)
- Mixture of Tokens: Continuous MoE through Cross-Example Aggregation (NeurIPS 2024 · similarity 0.87)
- FlexMoE: One-for-All Nested Intra-Expert Pruning for MoE Language Models (NeurIPS 2026 · similarity 0.87)
- Mixture-of-Experts Meets Instruction Tuning: A Winning Combination for Large Language Models (ICLR 2024 · similarity 0.87)
- Measuring and Guiding Monosemanticity (NeurIPS 2025 · similarity 0.86)
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.