Explaining the role of experts in personalization in MoE large language models

Polysemantic neurons in dense LLMs make personalization-related components difficult to locate, so the authors instead adapt MoEs to map different preferences to a few experts.

• Scenario: The authors consider personalization as adapting an LLM to user or group preferences, with demand coming from user-centric applications. Most existing methods modify dense LLMs, which the authors claim have little interpretability.
• Limitation: The authors claim that polysemantic neurons encode multiple unrelated concepts into the same neuron, hindering the localization of components encoding different personalization features, and also hindering understanding and control of personalization.
• Hypothesis: MoEs use sparse routing to send each token to a few experts, thereby using a wider, sparser network. The authors claim this reduces polysemanticity, allowing experts to encode different task features.
• Approach: Perform two-stage parameter-efficient fine-tuning on three open-source MoEs using group preference data, then use a crosscoder to compare personalized and non-personalized activations, identify personalization latents and their experts, and use causal interventions to check whether these components can change behavior.
• Problem decomposition: RQ1 compares polysemanticity and disentanglement under personalization settings; RQ2 identifies the separated personalization features; RQ3 locates the routing paths of different features.

In the introduction and §2, the authors divide related work into personalization, MoE routing, and feature and fine-tuning interpretability.

Personalization Methods

• Conditioning and parameter adaptation: Salemi et al. (2024), Richardson et al. (2023) condition on retrieved user history or profile summaries; Jiang et al. (2024b) encode persona in the prompt; Tan et al. (2024) train per-user adapters.
• Rewards, merging, and steering: Cheng et al. (2023), Jang et al. (2024) learn personalized reward models or merge preference policies; Zhang et al. (2025a), Cao et al. (2024) steer activations toward user style.
• Evaluation data: LaMP and LongLaMP evaluate per-user personalization; Multifaceted Collection and Bench use system messages to specify group preferences. The authors state these works treat dense models as black boxes.

MoE and Routing

• Architecture: Shazeer et al. (2017), Fedus et al. (2022), Zoph et al. (2022) route tokens to a few experts, decoupling parameter count from computation. Open models mentioned include Mixtral, OLMoE, Qwen3.
• Routing analysis: Jiang et al. (2024a), Xue et al. (2024) report specialization at the token and syntactic level rather than the domain level; Olson et al. (2025) study semantic routing in controlled settings; Bandarkar et al. (2026) identify language-specific experts using routing frequency comparisons.
• Exploiting specialization: Bai et al. (2025) tune the router to identify context-loyal experts; Li et al. (2026) encourage expert differentiation during training; Fayyaz et al. (2026) steer behavior by activating or deactivating experts. The authors adopt router-then-expert adaptation and link routing changes to persona features.

Feature Interpretability and Fine-Tuning Differences

• Polysemy and probing: Olah et al. (2020) discuss polysemy in dense neurons; Bricken et al. (2023), Huben et al. (2024) use SAEs to recover more monosemantic features; Gurnee et al. (2023) use sparse probing to measure how many neurons encode a concept.

• Sparse architectures: Yang et al. (2025b), Herbst et al. (2026) discuss how wider, sparser models (including MoE) reduce polysemy.

• Crosscoder: Minder et al. (2025) use a BatchTopK crosscoder to separate latents specific to fine-tuned models and apply it to chat tuning; Chaudhari et al. (2026) use it to compare MoE and dense models; Paulo et al. (2025) automatically interpret features. The authors subtract the personalized MoE from its base model and map the latents to the experts that route them.

• Baselines and benchmarks: Comparisons are dense LLMs matched in active or total parameter count, namely OLMo-2-1B/7B, Qwen2.5-3B/32B, Llama-2-13B/70B, and JANUS 7B fully fine-tuned on MULTIFACETED COLLECTION with ORPO/DPO. Training uses the Collection; evaluation uses MULTIFACETED BENCH, LaMP, LongLaMP.

The authors position this work beyond dense black-box personalization: studying how sparse MoE represents personalization and mapping routing changes after two-stage adaptation to interpretable persona features.

Overall, the approach is first two-stage PEFT, then use a crosscoder to decompose the activation difference of the personalized model relative to the base model into latents, and map them to experts. Preference P can be interaction history or a set of stated preferences; the generated writing is ŷ_p = M_θ(x, A_p(x, P)), scored relative to the gold reply. This paper implements A_p as PEFT parameters.

Two-stage adaptation

• Motivation: The authors claim that load balancing during pretraining distributes tokens uniformly across experts, which is not conducive to individual experts specializing in task features. Therefore, following Bai et al. (2025), they first find personalization-related experts, then tune only those experts.
• Router tuning: Only the router is updated, with experts and other parameters frozen, and the loss is the language modeling loss. Then the frequency with which the tuned router selects each expert on preference data is counted and divided by the routing k, so that the frequencies per layer sum to 1.
• Expert tuning: Only the k most-selected experts per layer are fine-tuned, with the router and other parameters frozen. The paper does not give the value of this k. In implementation, both MoE and dense use LoRA, and only MLP weights are updated.
• MoE forward pass: Each layer uses N_E experts and a router; tokens only enter the top-k experts; the output is the input residual plus the gated weighted transformation of the selected experts.

Crosscoder and personalization latents

• Training object: Paired activations from the same layer of the personalized and non-personalized models are collected on the same preference data to train a BatchTopK crosscoder. The expansion factor for all three MoEs is 32, with the sparsity term as L0; the appendix states activations are taken from residual MLP contributions. The paper does not report the BatchTopK k.
• Dictionary: A shared dictionary reconstructs both models simultaneously. Each latent has its own decoder vector for each model; if the decoder vector on the personalized side dominates, that latent mainly contributes to reconstructing the personalized model.
• Filtering: The relative decoder norm difference is Δ_j=1/2((|d^ft_j|₂−|d^base_j|₂)/(max(|d^ft_j|₂,|d^base_j|₂))+1), ranging from 0 to 1, with larger values more biased toward the personalized model. Δ_j ≥ 0.9 is assigned to the personalization set, ≤ 0.1 to the base set, and values in [0.4, 0.6] to the shared set.

Layer selection and expert mapping

• Layer selection: The authors claim middle layers carry the most information. Among them, they select the layer with the largest Jensen–Shannon divergence between personalized and non-personalized routing distributions, and require low routing entropy, meaning the router is confident in selecting experts.
• Mapping: For each personalization latent, take the N sequences with the highest activation; in these sequences, rank each layer's experts by "selection frequency on latent-activated tokens minus frequency on non-activated tokens," and keep the top-k′.
• Selecting (N, k′): Distinctiveness is 1 minus the average overlap of expert sets across latents, with higher meaning different latents correspond to different experts. Stability is the consistency of the expert set for the same latent across random subsamples of N sequences. The authors claim that too large an N causes sequences to overlap across latents, decreasing Distinctiveness. Both are grid-searched over [0, 1].

Explanation and causal operations

• Automatic explanation: Both the explainer and scorer are Llama-3.3-70B-Instruct. The explainer looks at 30 activation samples (top 20 plus 10 weaker ones), each a 40-token window centered on the maximum-activation token, with 32 before and 8 after. The scorer only sees the explanation and labels 20 activated and 20 non-activated samples.
• Sufficiency patching: At the crosscoder layer, replace the token's personalized activation with the non-personalized activation plus the decoder contribution of the personalization latent, then continue generation from subsequent layers, comparing the KL of the next-token distribution. The control is random latents matched in number and activation frequency, as well as shared-region latents.
• Routing steering: Collect experts aligned with the corresponding latent along the preference dimension, take the top-k′ by routing frequency, and add β times the standard deviation to the router logits of these experts on the non-personalized model, with β = 2. The control is the same number of randomly selected experts from the same layer.

Experimental Setup

• Tested MoEs: OLMoE-1B-7B (0125), Qwen3-30B-A3B (2507), Mixtral-8x7B (v0.1), all instruction-tuned decoder-only. Each reported with no adaptation, router tuning only (RT), and router plus expert tuning (RT+EFT).
• Controls: Dense models matching active and total parameters are OLMo-2-1B (0425)/OLMo-2-7B (1124), Qwen2.5-3B/Qwen2.5-32B, Llama-2-13B/Llama-2-70B, each with no adaptation and PEFT. Reference is the fully fine-tuned JANUS 7B and its ORPO/DPO variants. Table 2 writes Llama-2 as Llama-2-13b, Llama-2-70b.
• Data: Training and crosscoders use MULTIFACETED COLLECTION. Each item contains a query and system message, covering style, background knowledge, informativeness, harmlessness; aligned responses generated by GPT-4-Turbo-0125. Evaluation uses MULTIFACETED BENCH (same format, preferences disjoint from training set), LaMP, LongLaMP. The latter two infer per-user preferences from interaction history; following the user split of Zhang et al. (2025a), and omitting the email task whose data is not public. Figure 2 says it averages over 7 datasets.
• Metrics and generation: Report METEOR and ROUGE-L relative to gold (Table 2 records the latter as R). MULTIFACETED BENCH inference uses temperature 1.0, top-p 0.9, repetition penalty 1.03, max 4096 new tokens. The paper does not state sampling settings for LaMP and LongLaMP, nor report number of generation repetitions.
• Analysis protocol: k-sparse probing uses balanced accuracy; automatic interpretation uses F1 of Llama-3.3-70B-Instruct; causal comparison uses next-token KL. Figure 17 uses 921 MULTIFACETED BENCH samples. The paper does not report total sample counts for the Collection and each benchmark, per-layer k for expert tuning, LoRA rank, learning rate, training steps, or scorer-human agreement.

Main Results

The table below lists only RT+EFT MoEs, PEFT dense models with similar active parameters, and two JANUS variants. M is METEOR. Scholar, Tweet belong to LaMP, Review belongs to LongLaMP. According to Table 2.

Swipe horizontally to read the table

Model and variant	MFBench M	MFBench R	Scholar M	Tweet M	Review M
OLMoE-1B-7B RT+EFT	0.32	0.21	0.37	0.40	0.14
Qwen3-30B-A3B RT+EFT	0.37	0.26	0.41	0.39	0.23
Mixtral-8x7B RT+EFT	0.36	0.25	0.42	0.38	0.22
OLMo-2-1B PEFT	0.31	0.20	0.38	0.40	0.14
Qwen2.5-3B PEFT	0.35	0.24	0.39	0.38	0.21
Llama-2-13B PEFT	0.34	0.23	0.40	0.38	0.19
JANUS+DPO 7B	0.35	0.23	0.41	0.36	0.24
JANUS+ORPO 7B	0.21	0.10	0.09	0.29	0.07

• Authors' overall claim: The authors state that across families and metrics, MoEs outperform dense models with similar active parameters; for Qwen and Mixtral, Figure 2 averages match or exceed dense models with similar total parameters. In the table, the Scholar column already has cells where MoE is lower than the active control. On Table 2's Abstract Generation, Mixtral RT+EFT METEOR is 0.31, Llama-2-13B PEFT is 0.32.
• Which stage contributes gains: The authors state Qwen3's gains come mainly from router tuning, Mixtral more from expert tuning; gains on MULTIFACETED BENCH are larger than on LaMP and LongLaMP. In Table 2, Qwen3's RT+EFT is not always higher than RT: Scholar METEOR goes from 0.42 to 0.41, Tweet METEOR from 0.41 to 0.39.
• Parameters and reference: In §6.1 the authors state only 20% of total parameters are used per token, and the conclusion writes ∼20%. Table 1's active/total are OLMoE 1.3B/6.9B, Qwen3 3B/30B, Mixtral 13B/46.7B. Not listed in the table above are the no-adaptation and RT-only rows, the three dense groups matched on total parameters, and JANUS 7B. JANUS+ORPO has the lowest MFBench METEOR in Table 2 across the whole table; the paper does not explain why it is lower than JANUS 7B.

Polysemy and Disentanglement

• Probing setup: For each dimension, take the 5 most frequent sub-dimensions in the Collection, totaling 20 concepts. Positive examples are response tokens whose system message specifies that sub-dimension, with up to 1024 tokens per response and 100 queries; negative examples are other sub-dimensions of the same dimension, about 50,000 tokens per class. MoEs count only tokens routed to the target expert, and select the layer with highest accuracy on a 25% test split.
• Authors' claim about Figure 3: With similar active parameters, MoEs have higher balanced accuracy than dense models, with wider gaps within the same family (the authors write Olmo, Qwen), and the gap remains wide at k=1. The authors state this corresponds to monosemantic behavior for personalization. The text conversion does not preserve accuracy per k; bar heights are not transcribed.
• Sharing with the base: For personalization latents, compute ν = β_base/β_ft, where closer to 0 means less used to reconstruct the non-personalized model. The authors state that within the same family, MoEs have lower ν than dense models, with the exception that Mixtral is higher than Llama-13B with similar active parameters, and state this comparison is limited to the same model family. Figure 4's labels do not align with model pairs; bar values are not transcribed.

Content of Personalization Features

• Quantity: Personalization latents with Δ ≥ 0.9 are 157/65536 on Qwen3's L25, 132/131072 on Mixtral's L15, 111/65536 on OLMoE's L6. The authors state this is less than 0.25% of total latents. The selected layers come from the routing divergence and entropy criteria in §6.3.
• Activation location and dimensions: The authors state 85% of OLMoE's and 82% of Qwen3's activate mainly on the system message; Mixtral is an exception, with 41% activating mainly in the response. OLMoE is dominated by style 52% and harmlessness 21%; Qwen3 and Mixtral are dominated by background knowledge, 45% and 40% respectively.
• Interpretation quality: Overall F1 is OLMoE 0.65, Qwen3 0.63, Mixtral 0.70 (Figure 7). The authors state F1 increases monotonically with activation bins, with top bins at 0.81, 0.83, 0.88, in the same order as the three models above, and state interpretations best describe strong activations. Figure 8 and Table 3 give the highest-activation examples and top-expert token proportions; the preference quotes therein are not transcribed.

Routing Stability and Causal Intervention

• Qwen3 mapping: The authors take N=400, 1 expert per layer, and report D=0.94, S=0.81, and state samples of different latents are routed to their respective experts. OLMoE and Mixtral grids are in Figures 15, 16; the main text does not specify which cell is finally used.
• Patching: The authors state that on MULTIFACETED BENCH, personalization latents yield larger KL decreases than random and shared controls (Figure 10 top). Figure 17 repeats this comparison on the first 9 response tokens and 921 samples, with controls as the mean of 3 samplings, p<0.001 relative to both controls (Wilcoxon, Holm correction), and gives 95% bootstrap confidence intervals. The main text does not give the KL decrease in percentage points.
• Steering: For each dimension, steer toward the top-k′ experts aligned with that dimension's latent, β=2. The authors state that on each dimension, this steering has lower KL than random steering (Figure 10 bottom, p<0.001 relative to controls), and state this corresponds to different routing for different personalization features. The caption only defines the marker for p<0.001.

Other Ablations and Analyses

• Qwen3's D (Figure 9, k′=1/2/4/8): 0.94, 0.94, 0.92, 0.89 at N=400; 0.95 at N=100, k′=1.
• Qwen3's S: 0.81, 0.78, 0.78, 0.79 at N=400 for k′=1/2/4/8; 0.83 at N=500, k′=1.
• OLMoE (Figure 15): D at N=400 for k′=1/2/4/8 is 0.90, 0.89, 0.86, 0.80; S in the same cell at k′=1 is 0.95.
• Mixtral (Figure 16, only k′=1, 2): D is 0.79 for N=100 to 500, k′=1; S is 0.94 at N=500, k′=1.
• Variance explained by reconstruction: OLMoE 0.80, Qwen3-30B-A3B 0.74, Mixtral 0.65 (Limitations).
• Cells in Table 2 that decrease after personalization: OLMoE's Review METEOR goes from 0.21 without adaptation to 0.14 with RT+EFT. The paper does not explain this.

Limitations and future directions noted by the authors

• Models and layers: Only three open MoE families are studied, and for each model only one layer selected by routing divergence is analyzed. The authors say personalization features may also appear in other layers, and conclusions may not transfer to other designs such as shared experts or fine-grained expert splitting (Limitations, Models and comparisons).
• Comparisons cannot be fully separated: Dense baselines are matched by active or total parameter count, but pretraining data and recipes differ. The authors say architectural effects cannot be fully separated from training differences (same section).
• Definition of preference: Multifaceted-Collection uses synthetic system messages across four dimensions, with sub-dimensions that are categorical and long-tailed. The authors say this captures group preferences rather than individual users, experiments are limited to English, and real user histories may be noisier and more entangled (Data and notion of persona).
• Crosscoder: Reconstruction is imperfect, with variance explained of OLMoE 0.80, Qwen3-30B-A3B 0.74, and Mixtral 0.65. The authors say information in reconstruction error is invisible to latent analysis, and personalization latents depend on a decoder norm threshold; BatchTopK and latent scaling reduce sparse artifacts but cannot fully rule them out (Crosscoder fidelity).
• Probing: Persona is a property of the whole response, but sparse probing labels individual tokens, which the authors say introduces label noise and lowers absolute scores. Selecting the best component across layer × expert favors MoE; the authors say selection was done on the test set and test scores are reported, but selection effects may remain, especially at k=1. The probe measures whether a concept can be decoded, not whether the model uses it (Probing).
• Attribution and generation metrics: Expert sets come from routing frequency of firing tokens, which the authors call correlational, depending on sample count, selection rule, and background routing rate, with moderate stability. ROUGE-L and METEOR measure overlap with reference text, which the authors say does not proxy personalization well; sampling adds variance; automatic explanations rely on Llama-3.3-70B-Instruct, whose biases may affect descriptions (Expert attribution and causality; Evaluation of generation).

Experimental coverage

• Models: Tested MoEs are OLMoE-1B-7B (0125), Qwen3-30B-A3B (2507), and Mixtral-8x7B (v0.1). Baselines are OLMo-2, Qwen2.5, and Llama-2 at two sizes each from Table 1, plus JANUS 7B and ORPO/DPO (§5, Table 2).
• Benchmarks: Scores are reported on MULTIFACETED BENCH, LaMP's News/Scholar/Tweet, and LongLaMP's Abstract/Topic/Review, with metrics METEOR and ROUGE-L. The email task is omitted because the data are not public (§5).
• Analysis: k-sparse probing uses 20 concepts, about 50,000 tokens per class, and a 25% test split. Crosscoder and routing analyses each take one layer: Qwen3 L25, Mixtral L15, OLMoE L6. Both explainer and scorer are Llama-3.3-70B-Instruct (§6.2–§6.3).
• Interventions: Include activation patching and expert steering along the four preference dimensions. Figure 10 notes Wilcoxon, Holm correction, p<0.001. Figure 17 uses 921 MULTIFACETED BENCH samples, 3 control samplings, and 95% bootstrap confidence intervals.
• Not reported: The paper does not report the per-layer expert count k for expert tuning, the k for BatchTopK, LoRA rank, learning rate, number of training steps, number of generation repetitions, or agreement between the LLM scorer and human judgment.

The authors analyze what each expert encodes during MoE personalization: first adapting to group preferences, then decomposing the activation difference before and after adaptation.

• Problem: Dense neurons encode multiple unrelated concepts simultaneously, making personalization-related components hard to locate. The authors instead use MoE sparse routing, so each token passes through only a few experts.
• Method: First tune only the router, then tune only the most-selected experts per layer. Then train a BatchTopK crosscoder on layers with the largest routing distribution difference and low routing entropy, select personalization latents with decoder relative norm difference not below 0.9, and map them to experts by the routing frequency difference between activated and non-activated tokens.
• Generation scores: In Table 2, RT+EFT Qwen3-30B-A3B achieves METEOR 0.37 and ROUGE 0.26 on MULTIFACETED BENCH; Mixtral-8x7B achieves 0.36 and 0.25; OLMoE-1B-7B achieves 0.32 and 0.21. The authors claim this is comparable to or better than PEFT dense with similar active parameters, but the table still has lower cells. The authors claim only 20% of total parameters are used per token, and the conclusion writes ∼20%; Table 1 shows active/total as 1.3B/6.9B, 3B/30B, 13B/46.7B.
• Disentanglement: Personalization latents are 157/65536 for Qwen3, 132/131072 for Mixtral, 111/65536 for OLMoE. OLMoE and Qwen3 have 85% and 82% primarily activating on the system message, respectively; Mixtral has 41% primarily activating in the reply. Overall explanation F1 is 0.65, 0.63, 0.70. The authors claim that within the same family, MoE personalization latents are less shared with the base model, with the exception that Mixtral's ν is higher than Llama-13B.
• Routing: Qwen3 at N=400 and 1 expert per layer has D 0.94 and S 0.81. The authors claim that patching these latents, or boosting the corresponding experts on the non-personalized model, both reduce the next-token KL to the personalized model, with p<0.001 relative to the control.
• Authors' conclusion: MoEs encode different personalization features into different routes with lower polysemanticity, so they can be located and intervened upon. The authors also place the analysis on one layer per model, English, and group-level synthetic preferences, and state that overlap with reference text cannot be equated with preferences being satisfied.