SAE Interprets MoE: Sparse AutoEncoders Reveal That MoE Experts Exhibit Human-Interpretable Fine-Grained Specialization
Abstract
A sparse autoencoder trained on mixture-of-experts layer outputs reveals that individual experts drive specific latent features, showing finer-grained and more concentrated specialization than routing or domain labels alone.
Mixture-of-Experts (MoE) models route each token to a few of many experts, yet what an individual expert is responsible for remains unsettled. Existing analyses count how often an expert is selected on domain-labeled corpora, an instrument that fixes the taxonomy in advance, stays coarse, and records selection rather than contribution. We instead study specialization in feature space by training a sparse autoencoder (SAE) on the output of an MoE layer. Since this output is a router-weighted sum of expert outputs and the SAE encoder is affine, the pre-activation of every latent decomposes exactly over the experts. Aggregated over a corpus, this gives each expert a dominant set of latents whose activation it principally drives. On layers 4, 30 and 46 of Qwen3-30B-A3B, contribution is far more concentrated than gating, most latents have only one or two dominators, and the overlap between dominant sets concentrates on a few expert pairs. Auto-interpreting these sets reveals a division of labor below coarse domains, with distinct experts for Java and front-end web code, or for geometry and for arithmetic and polynomials, while the coarse domains themselves take very unequal numbers of experts. Our SAEs and code are available at https://github.com/Bytedance-Seed/SAE-Interprets-MoE.
Community
TL;DR: We train a TopK SAE with 2^20 latents and k = 100 on the MoE block output of Qwen3-30B-A3B at layers 4, 30 and 46. Because the SAE encoder is affine and the MoE output is a weighted sum of expert outputs, the activation of each latent decomposes exactly over the activated experts. Aggregating this decomposition over 1.5B tokens gives every expert a dominant set of latents that it principally drives. This lets us read what each expert does without any predefined domain taxonomy.
Key findings:
- Contribution is far more concentrated than gating.
- Most latents have only one or two dominators. Dominant sets are largely private to each expert.
- Experts specialize below coarse domains. A Java expert sits next to a separate expert for front end web code, and a geometry expert sits next to one for arithmetic and polynomials.
- Grouping the 128 experts of layer 30 gives 11 domains of very unequal size. Some domains are not subject areas at all, such as text form, language, and entities and identifiers.
Code, the three SAEs, and an interactive explorer of all 128 experts of layer 30 are released. Happy to answer questions!
Models citing this paper 1
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper