Personalization of Large Language Models (LLMs) is increasingly important for user-centric applications, yet prior methods mainly adapt dense LLMs, whose polysemantic neurons, each encoding multiple unrelated concepts, hinder localizing the components behind personalization, and thus understanding and controlling it. We propose to personalize Mixture-of-Experts (MoE) LLMs, which sparsely route each token to a few specialized experts, reducing polysemanticity and enabling such localization. We adapt three open-weight MoE LLMs on a large-scale personalization dataset and train crosscoders on paired activations of the personalized and non-personalized models to learn interpretable personalized latents, from which we identify the experts encoding distinct personalization features. The personalized MoE LLMs perform competitively with dense LLMs while activating only a fraction of their parameters, exhibit lower polysemanticity than dense LLMs with similar active parameters, and, within the same model family, yield personalized latents that are less shared with the non-personalized model. Causal interventions show that these latents suffice to recover the personalized model's behavior and that boosting their associated experts moves the non-personalized model toward it, revealing distinct routing paths for different personalization features.
Figures & tables
Figure 1: Overview of the proposed framework: [Bottom → Top] For a given query and user preference via system message, we collect activations for all tokens of the input prompt from a particular layer of personalized (Left) and non-personalized (Right) MoE and then train a SAE-based model (crosscoder) to learn and identify a set of latents associated with the personalized MoE LLMs capturing distinct personalization features.
Figure 2: Personalization performance of LLMs: each box plot shows average performance over personalized and non-personalized variants on 7 datasets (points), with the mean performance represented using ⋄ . The detailed result is described in Table 2 .
Figure 3: K-sparse probe balanced accuracy comparison between MoE and dense LLMs with similar active parameters.
Figure 4: Ratio ν of scales to explain non-personalized and personalized models via personalized latents and their behavior across dense and MoE LLMs.
Figure 5: Layer-wise routing distribution analysis. Top: routing JS-divergence between the personalized and non-personalized Qwen3-30B-A3B (2507). Bottom: entropy of the two models with overlapped experts (right axis). L25 is identified as the best layer. Refer to Figures 11 and 12 for other models.
Figure 6: Relative norm difference distribution of latents from the crosscoder trained on activations from L25 of personalized and non-personalized Qwen3-30B-A3B (2507). Refer Figures 13 and 14 for other models.
Figure 7: F1 for classifying positive and negative examples activated by latents (Overall). We also show its efficiency across different activation quantiles.
Figure 8: Top activating latent from Qwen3-30B-A3B (2507): its autointerpretability explanation, dominant preference dimension in MULTIFACETED COLLECTION, the top- k′ expert at layer 25 with the share of the latent’s activated tokens routed to it, and an excerpt from a top- N activating sequence with the maximally activating token highlighted. More top activating latents from different models are described in Table 3 .
Figure 9: Grid search over (N,k) for Qwen3-30B-A3B (2507). Top: Distinctiveness of top- k′ experts across latents with varying N . Bottom: Stability of top- k′ experts for the same N examples under resampling for a particular latent.
Figure 10: Top: KL patching at the crosscoder layer over full responses. ‘Random’ and ‘Shared’ baselines are shown as the mean of 3 draws with dots. Bottom: Steering top- k′ experts per dimension in the non-personalized model; measuring KL divergence with its personalized counterpart. ***: statistically significant with p<0.001 vs. controls (Wilcoxon, Holm-corrected).
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Active
Total
Params
Params
OLMoE family
OLMo-2-1B (0425)
1B
1B
OLMoE-1B-7B (0125)
1.3B
6.9B
OLMo-2-7B (1124)
7B
7B
Qwen family
Appendix
Table 1: For each MoE LLM, we compare with similar active and total parameter-sized dense LLMs with the same model family if available. For dense LLMs, active parameters are the same as total parameters.
Model
Act.
MFBench
LaMP
LongLaMP
News
Scholar
Tweet
Abstr.
Topic
Review
M
R
M
R
M
R
M
R
M
R
M
R
M
R
OLMo-2-1B (0425)
1B
0.29
0.19
0.12
0.08
0.30
0.18
0.29
0.22
0.31
0.18
0.23
0.11
0.18
0.14
OLMo-2-1B (0425) ⋆
1B
0.31
0.20
0.12
0.12
0.38
0.29
0.40
0.38
0.26
0.19
0.23
0.12
0.14
0.12
OLMoE-1B-7B (0125)
1.3B
0.25
0.20
0.12
0.11
0.36
0.28
0.35
0.32
0.31
0.18
0.24
0.11
0.21
0.14
OLMoE-1B-7B (0125) †
1.3B
0.32
0.20
0.12
0.12
0.35
0.26
0.36
0.32
0.31
0.17
0.23
0.11
0.19
0.14
Appendix
Table 2: Personalization performance comparison over LLMs across personalization benchmarks using two evaluation metrics METEOR (M) and ROUGE (R). Act. indicates active parameters of LLMs. ⋆ indicates PEFT fine-tuned (FT) dense LLMs, † indicates router-tuned (RT) MoE LLMs, and ‡ indicates the two stage router and expert (RT + EFT)-tuned MoE LLMs. Row colors: dense, active size , MoE , dense, total size . Acronym of personalization benchmarks: MFBench = MULTIFACETED BENCH; LaMP : News = News Headline Generation, Scholar = Scholarly Title Generation, Tweet = Tweet Paraphrasing; LongLaMP : Abstr. = Abstract Generation, Topic = Topic Writing, Review = Review Generation. Within each model family (dense and MoE models together), metric and benchmark, bold indicates the best and underline indicates the second-best result, and tied values share the mark.
Figure 11: Layer-wise routing distribution analysis. Top: routing JS-divergence between the personalized and non-personalized OLMoE-1B-7B (0125. Bottom: entropy of the two models with overlapped experts (right axis). L6 is identified as the best layer.
Figure 12: Layer-wise routing distribution analysis. Top: routing JS-divergence between the personalized and non-personalized Mixtral-8x7B (v0.1). Bottom: entropy of the two models with overlapped experts (right axis). L15 is identified as the best layer.
Figure 13: Relative norm difference distribution of latents from the crosscoder trained on activations from L6 of personalized and non-personalized OLMoE-1B-7B (0125).
Figure 14: Relative norm difference distribution of latents from the crosscoder trained on activations from L15 of personalized and non-personalized Mixtral-8x7B (v0.1).
Model
Latent
Explanation
Dimension
Top- k′ Experts
Example
OLMoE-1B-7B (0125)
12503
The feature detects the presence of meticulously structured and precise language, often characterized by the use of advanced technical terminology, detailed explanations, and a formal tone, typically in the context of scientific, technical, or academic discussions.
informativeness
L6-E9 (0.91);L6-E22 (0.73)
“…Your discussions are met iculously detailed, exploring biochemical mechanisms of Vitamin C…”
Mixtral-8x7B (v0.1)
45358
The feature detects tokens related to grammatical correctness, punctuation, and sentence structure, often in the context of editing, correcting, or explaining language usage.
style
L15-E2 (0.96);L15-E5 (0.34)
“…identifying run-on sentences with a focus on gram mat ical terms, catering to an audience with prior knowledge…”
Appendix
Table 3: Examples of personalized latents identified by the crosscoder for three MoE LLMs: autointerpretability explanation, dominant preference dimension from the MULTIFACETED COLLECTION benchmark, top- k′ experts at the optimal layer with the percentage of activated token routed towards that expert, and a representative example from top- N maximally activating sequences with the maximally activated token highlighted.
Figure 15: Grid search over (N,k) for OLMoE-1B-7B (0125). Top: Distinctiveness of top- k′ experts across latents with varying N . Bottom: Stability of top- k′ experts for the same N examples under resampling for a particular latent.
Figure 16: Grid search over (N,k) for Mixtral-8x7B (v0.1). Top: Distinctiveness of top- k′ experts across latents with varying N . Bottom: Stability of top- k′ experts for the same N examples under resampling for a particular latent.
Figure 17: KL patching at the crosscoder layer over the first 9 response tokens. Reduction of the next-token KL divergence to the personalized MoE, relative to patching in the non-personalized MoE output, when adding the decoder contributions of the personalized latents (ft-only) or of equally sized, firing-frequency-matched random or shared latents (mean of 3 draws; dots: individual draws). Error bars: 95% bootstrap CIs over the 921 Multifaceted Bench examples; ***: p<0.001 vs. both controls (Wilcoxon, Holm-corrected).
Large language models (LLMs) are increasingly expected to adapt to individual users, yet effective personalization remains challenging when only limited user-specific samples are available. In this work, we take an alternative perspective: pretrained LLMs may already possess latent capacity for personalization, and a few user samples may therefore suffice to guide the model toward user-aligned behavior with minimal user-specific adaptation. From this perspective, we propose LatentPersonal, a framework that formulates personalization as navigation in a shared latent adaptation space. LatentPersonal infers a compact latent representation from a few user samples to guide user-specific model adaptation, regularized with a variational information bottleneck to encourage compact preference representations. We instantiate LatentPersonal with LoRA, leveraging its low-rank parameterization as a natural low-dimensional adaptation space for personalization. By simply inserting a user-specific guidance vector between the shared low-rank factors, the model can navigate toward personalized adaptations through lightweight inference of this compact representation, without updating the shared LoRA parameters. Experiments across multiple personalization datasets demonstrate that LatentPersonal substantially reduces user-specific adaptation overhead while achieving effective personalization from only a few user-specific interactions, with particularly strong performance in the one-shot regime.
Wei Chen, Guanghui Zhu, Zhongliang Cai +1
State Key Laboratory for Novel Software Technology, Nanjing University
Large Language Models (LLMs) exhibit strong implicit personalization ability, yet most existing approaches treat this behavior as a black box, relying on prompt engineering or fine tuning on user data. In this work, we adopt a mechanistic interpretability perspective and hypothesize the existence of a sparse set of Preference Heads, attention heads that encode user specific stylistic and topical preferences and exert a causal influence on generation. We introduce Differential Preference Steering (DPS), a training free framework that (1) identifies Preference Heads through causal masking analysis and (2) leverages them for controllable and interpretable personalization at inference time. DPS computes a Preference Contribution Score (PCS) for each attention head, directly measuring its causal impact on user aligned outputs. During decoding, we contrast model predictions with and without Preference Heads, amplifying the difference between personalized and generic logits to selectively strengthen preference aligned continuations. Experiments on widely used personalization benchmarks across multiple LLMs demonstrate consistent gains in personalization fidelity while preserving content coherence and low computational overhead. Beyond empirical improvements, DPS provides a mechanistic explanation of where and how personalization emerges within transformer architectures. Our implementation is publicly available.
Weixu Zhang, Ye Yuan, Changjiang Han +7
McGill University · Mila - Quebec AI Institute · MBZUAI +2
Proprietary Large Language Models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet aligning their outputs with diverse user preferences remains challenging. Existing personalization approaches for black-box LLMs often rely on user-specific scoring heads, causing the number of personalized parameters to grow linearly with the number of users and requiring additional adaptation for unseen users. To address these limitations, we propose Mixture-of-Facets (MoF), a scalable personalization framework for black-box LLMs that models user preferences as compositions of shared latent preference facets rather than dedicated user-specific parameters. MoF performs personalization through history-conditioned routing over shared facet heads, enabling personalization for users unseen during training without additional parameter updates. Across diverse personalization tasks, MoF delivers stronger personalization performance while maintaining a more scalable and parameter-efficient design than prior approaches. Additional analysis indicates strong generalization to unseen users.
Hun Park
Department of Artificial Intelligence Korea University Seoul, Republic of Korea