RouterInterp: Understanding Superposed Specialisation in Mixture of Experts Routing
cs.AI, cs.CL, cs.LG
Submitted: 2026-10-08
Updated: 2026-10-08
Comments: 33 pages (12 non-appendix pages), 7 figures, published as a conference paper at ICML 2026
Code: https://github.com/ilyalasy/routerinterp
Project page: https://gabgoh.github.io/ThoughtVectors
License: http://creativecommons.org/licenses/by/4.0/
The gist: Sparse Mixture of Experts (MoE) models scale more efficiently than dense models by routing tokens to modular expert networks that are only active for processing a fraction of tokens.
Terminology
Abstract
Sparse Mixture of Experts (MoE) models scale more efficiently than dense models by routing tokens to modular expert networks that are only active for processing a fraction of tokens. A leading hypothesis for the performance of MoE models is that each expert specialises in a single, coherent domain. However, interpretability efforts that assume this hypothesis have generally been unsuccessful. We propose and present evidence for an alternative account that we call the Superposed Specialisation Hypothesis (SSH): experts specialise in a disjoint union of fine-grained features rather than one broad domain. Leveraging the SSH, we introduce RouterInterp, a method for interpreting expert routing that identifies Sparse Autoencoder features most predictive of routing decisions and produces unified natural language explanations. On gpt-oss-20b, RouterInterp explains expert routing with about 65% higher detection accuracy than prior token statistics based methods. This work provides a scalable method for generating more accurate explanations of expert routing and increases our understanding of a previously uninterpretable component of foundation models.
Sources
- On the Complexity of Neural Computation in Superposition
- gpt-oss-120b & gpt-oss-20b Model Card
- Eliciting Latent Predictions from Transformers with the Tuned Lens
- Do Sparse Autoencoders Capture Concept Manifolds?
- The Local Interaction Basis: Identifying Computationally-Relevant and Sparsely Interacting Features in Neural Networks
- A Review of Sparse Expert Models in Deep Learning
- Clusterability in Neural Networks
- The Pile: An 800GB Dataset of Diverse Text for Language Modeling
- Adaptive Computation Time for Recurrent Neural Networks
- Mixtral of Experts
- Interpretable Embeddings with Sparse Autoencoders: A Data Analysis Toolkit
- Scaling Laws for Fine-Grained Mixture of Experts
- DeepSeek-V3 Technical Report
- k-Sparse Autoencoders
- Understanding polysemanticity in neural networks through coding theory
- Efficient Estimation of Word Representations in Vector Space
- Sparse Autoencoders for Hypothesis Generation
- Mixture-of-Depths: Dynamically allocating compute in transformer-based language models
- AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
- Qwen3 Technical Report
Related papers
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Agents' Last Exam
- The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
- MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
- Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems