Mixture-of-Experts layers can carry a hundred or more experts, and the standing question has always been blunt: what does any single one of them actually do? The default answer counts selection frequency — run domain-labeled corpora through the model and tally how often the router picks each expert. A ByteDance-Seed paper posted to arXiv on November 5, "SAE Interprets MoE", swaps the ruler entirely: instead of counting who gets picked, it computes who actually contributes activation. On Qwen3-30B-A3B the resulting picture is unusually concrete — Java code and front-end web code belong to different experts, and geometry splits from arithmetic and polynomials. The division of labor inside MoE turns out to be far finer than the usual "an expert for code, an expert for math" story.
Three blind spots in the old method
The paper opens by dismantling selection-frequency analysis on three counts. First, the taxonomy is fixed in advance: you decide which domains the corpora are labeled with before looking, so the experts never get to speak for themselves. Second, the granularity is coarse — at best you learn that an expert leans toward code. Third, and most damaging, it records the router's selection, not the expert's contribution. A routing weight is a routing weight; it is not a receipt proving the expert moved the output by the same share.
The free lunch from linear algebra
The core of the method is one very clean piece of math. A MoE layer's output is the router-weighted sum of the activated experts' outputs, and the encoder of a sparse autoencoder is an affine map. Stack the two and the pre-activation of every SAE latent decomposes exactly over the experts — not an approximation, not a regression, but an identity that holds algebraically. Aggregate those contribution fractions over a corpus and threshold them, and you get a bipartite dominance graph between experts and latents: each latent is dominated by a few experts, and auto-interpreting the latents an expert dominates yields that expert's job description. The SAE here is not just a reading aid — it doubles as an exact contribution-accounting machine.
What they found on Qwen3-30B-A3B
The team trained SAEs and ran the decomposition on layers 4, 30 and 46 of Qwen3-30B-A3B. Three findings stand out:
- Contribution is far more concentrated than gating: most latents have only one or two dominant experts.
- Overlap is localized too: the overlap between dominant sets concentrates on a few expert pairs rather than smearing uniformly.
- Specialization sits below coarse domains: within code, Java and front-end web code get distinct experts; within math, geometry separates from arithmetic and polynomials. Meanwhile the coarse domains themselves take very unequal numbers of experts — some domains hold a crowd, others barely any.
The auto-interpretation step used GPT-5.5 (the repo pins gpt-5.5-2026-04-24). All 128 experts of layer 30 were labeled and partitioned into 11 domains; the explanation texts, labels and theme judgements ship with the code, and the project page includes an interactive explorer where you can click through each of the 128 experts, its word clouds and its responsibilities.
Reproduction cost, and the openness that comes with it
Replication has a hardware cover charge: Qwen3-30B-A3B in bf16 takes about 61 GB, each traced layer's fp32 SAE encoder adds roughly 8 GB, and tracing all three layers in one pass needs two 80 GB GPUs, with transformers pinned below version 5. The statistics rest on a roughly 1.5-billion-token validation set drawn from the Olmo 3 training corpus (5/8 Dolma 3 pretraining data, 1/4 Dolci-Think, 1/8 Dolci-Instruct), kept disjoint from the data the SAEs were trained on. The SAEs themselves are TopK — 2^20 latents (about 1.05 million) with k = 100 — and the weights for all three layers are on Hugging Face under ByteDance-Seed/Qwen3-30B-A3B-MoE-SAE, with a huggingface.cn mirror for good measure. The code is Apache 2.0, and every experiment directory carries a run.sh that reproduces the paper's settings. For any team wanting the same expert-division map on their own model, the barrier has dropped from "invent the method" to "borrow two 80 GB cards".
So what
The value here goes beyond "one more model interpreted". It moves MoE interpretability from selection statistics to exact contribution accounting, and it hands downstream engineering a practical instrument. This site recently covered a MoE pruning study: pruning experts calibrated on English cost Lithuanian 21.4 points while English lost only 1.7. Why so lopsided? Because nobody actually knew which expert was responsible for what. With a reliable division-of-labor map in hand, pruning, router debugging and load-balancing decisions finally have something to stand on. The black box was never sealed; mostly, we were holding the wrong key.
References: arXiv paper · GitHub repository (code, data, SAEs)