Sparse Autoencoders Interpret Mixture-of-Experts Models by Revealing That Their Experts Exhibit Human-Interpretable Fine-Grained Specialization
This release uses sparse autoencoders to show Qwen3 MoE experts have private, fine-grained specializations, releasing code and an expert explorer.
Key Points Summary
- Researchers trained TopK sparse autoencoders with 2^20 latents and k = 100 on Qwen3-30B-A3B MoE outputs at layers 4, 30, and 46 to reveal what each expert does without predefined labels.
- The exact attribution works because the SAE encoder is affine and the MoE output is a weighted sum of expert outputs, allowing latent activations to be decomposed over activated experts using 1.5B tokens.
- Experts show highly concentrated and mostly private specialization: most latents have only one or two dominant experts, and these dominant sets are more focused than the gating/routing mechanism.
- Specialization is finer than broad subject domains, with separate experts for tasks such as Java versus front-end web code and geometry versus arithmetic and polynomials, including non-subject patterns like text form, language, and entities.
- The release includes code, three SAEs, and an interactive explorer for all 128 layer-30 experts, supporting the idea that MoE models organize knowledge into fine-grained, expert-owned patterns rather than clean human-defined categories.
For readers interested in model interpretability, a release describes training a TopK sparse autoencoder (SAE), a tool that extracts sparse internal features from model activations, with 2^20 latents and k = 100 on the MoE block output of Qwen3-30B-A3B at layers 4, 30 and 46. MoE, or mixture-of-experts, models split computation among many expert blocks. The release shows how to read what each Qwen3 MoE expert does without predefined domain labels.

Exact attribution across experts
The SAE encoder is affine, and the MoE output is a weighted sum of expert outputs. Because of that structure, 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.
Specialists, not broad domains
The analysis finds that contribution is far more concentrated than gating. Gating is the routing mechanism that decides which experts are used for an input. Most latents have only one or two dominators, and 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.
Released for inspection
The release includes code, the three SAEs, and an interactive explorer of all 128 experts of layer 30. The release says questions are welcome.
For readers following model interpretability, the value is concrete: it offers a way to connect expert routing with sparse features, and it suggests that MoE models may organize knowledge into fine-grained, expert-owned patterns rather than clean, human-defined categories.
Comments (0)