TL;DR of the paper, generated automatically.

A new paper uses a sparse autoencoder to show what each MoE expert actually does.

• Method: trains a sparse autoencoder on MoE layer outputs, not router choices, across three layers of Qwen3-30B-A3B.
• Finding: each expert's contribution is more concentrated than its gating, and experts specialize below coarse domains.
• Examples: separate experts for Java and front-end web code, and for geometry versus arithmetic and polynomials.