SAE Interprets MoE: Sparse AutoEncoders Reveal That MoE Experts Exhibit Human-Interpretable Fine-Grained Specialization
Mixture-of-Experts (MoE) language models contain many expert networks, but a basic question remains difficult to answer: what does any one expert actually do? A router selects a small number of experts for each token, yet selection frequency does not necessarily reveal functional responsibility. An expert may be selected often while contributing little to the representation produced by the layer.
In this paper, Anyuan Zhuo, Shaohua Zhang, Yuan Lin, and Pinyan Lu propose studying specialization in the feature space of an MoE layer. Their method combines MoE contribution analysis with sparse autoencoders (SAEs), which learn sparse, potentially interpretable features from model activations. The authors argue that this avoids imposing a human taxonomy such as “code” or “math” before examining how the model organizes its work.
From routing decisions to feature contributions
An MoE layer produces its output by taking a router-weighted sum of the outputs from the selected experts. The researchers train an SAE on that combined output. Because the SAE encoder is affine, the pre-activation of each SAE latent can be decomposed algebraically into an expert-specific contribution plus the encoder bias.
For each token prefix, expert, and active latent, the authors calculate the positive contribution made by the expert. They then normalize these contributions and average them over a large validation corpus. This produces a contribution fraction describing how much an expert generally accounts for a latent’s activation.
The authors call an expert a “dominator” of a latent when its contribution exceeds a chosen threshold. Connecting experts to the latents they dominate creates a dominance graph. An expert’s dominant set can then be interpreted as a description of the features it chiefly helps produce.
Evidence from Qwen3-30B-A3B
The study examines layers 4, 30, and 46 of Qwen3-30B-A3B. Each studied layer has 128 experts, with eight selected for each token. The researchers train a separate SAE for each layer, using 220 latent dimensions, and compute statistics on a 1.5-billion-token validation set.
The first result is that contribution is substantially more concentrated than gating. At layer 30, for example, the first-ranked expert accounts for an average contribution share of 0.512 for an active latent, compared with a gating share of 0.231. By the eighth rank, the average contribution share is 0.005, while the corresponding gating share is 0.074. The same qualitative pattern appears at layers 4 and 46.
This suggests that being selected is not equivalent to being important. Although eight experts participate in a token’s routing decision, a given latent is usually driven mainly by one or two of them. With a dominance threshold of 0.2, more than half of the latents in every studied layer have exactly one dominator. Including latents with two dominators covers between 85.1% and 92.7% of latents across the three layers.
The expert-side results point to distributed specialization rather than a few dominant hubs. Every expert dominates some nonempty set of latents, and the ten experts with the largest dominant sets account for no more than 30.1% of all dominance edges. Pairwise overlap is generally tiny: between 37.5% and 40.5% of expert pairs share no dominant latent at all. The overlap that does exist is concentrated in a small number of expert pairs.
Fine-grained roles inside broad domains
To interpret dominant sets, the authors use GPT-5.5 to summarize latent explanations based on activating contexts and logit effects. This generated labeling is evidence about the patterns surfaced by the method, not an independent human annotation study.
The resulting examples indicate specialization below conventional domain labels. In layer 30, expert 50 is associated primarily with Java and Android development, while expert 57 is associated with front-end web development involving HTML, CSS, and JavaScript. Within mathematics, expert 1 is linked to geometry, whereas expert 76 is associated with arithmetic and polynomial computation.
The paper reports that 86.7% of sampled latent explanations from expert 50 were judged related to Java development, compared with 1.9% in a sample from other experts. The corresponding shares were 78.6% for front-end development and 78.1% for geometry. These judgments support the claim that dominant sets are more informative than raw routed-token counts, though they depend partly on an automated interpreter.
When the 128 layer-30 experts are grouped into broader domains, the allocation is uneven. Computing contains 28 experts and mathematics and reasoning contains 20. Text form, language, and entities and identifiers together account for 36 experts, showing that the learned organization includes formatting, naming, and linguistic functions as well as subject matter.
Limitations and practical implications
The authors identify several important limitations. SAE reconstruction is imperfect, and learned features can be split across multiple latents or absorbed into others. The study analyzes one model, so the results do not establish that all MoE systems specialize in the same way. Its statistics also come from a researcher-chosen corpus rather than Qwen3’s undisclosed pretraining mixture.
The proposed link between expert allocation and the composition of pretraining data is therefore a conjecture, not a demonstrated causal explanation. Similarly, automated latent explanations and expert labels can introduce interpretation errors, even though the paper provides prompts and confidence procedures.
Practically, the method offers a more functional lens for MoE observability. Researchers could use dominance graphs to compare layers, routing objectives, or load-balancing strategies. The authors also suggest that dominant sets may provide testable guidance for expert pruning, merging, or feature-level steering: changing an expert should chiefly affect the latents it dominates. That possibility remains future work, however. The paper demonstrates attribution structure, not successful pruning or controlled editing.
The central contribution is methodological. By separating routing from contribution and analyzing learned features rather than fixed domains, the authors show that MoE experts can exhibit recognizable, fine-grained division of labor. The result is a promising map of specialization—but one that still needs validation across models, datasets, SAEs, and causal interventions.
Sources
AI-assisted research summary. Check the original paper for complete methods, results, and limitations.