Source-linked AI summary
Explaining the Roles of Experts in Mixture-of-Experts LLMs for Personalization
Sekh Mainul Islam, Pepa Atanasova, Isabelle Augenstein
TL;DR
Dense LLMs pose an interpretability challenge for personalization because polysemantic neurons make it difficult to localize the components encoding preferences. This paper adapts MoE models and uses crosscoders to identify personalized latents and their expert mappings; the models perform competitively with dense baselines while using a fraction of their parameters, and interventions support distinct personalization pathways.
Problem
Dense LLMs are difficult to interpret for personalization because polysemantic neurons hinder localization of components encoding distinct personalized features.
Source: §1 · p. 1Method
The paper adapts MoE LLMs with two-stage fine-tuning, trains crosscoders on paired personalized and base-model activations, and maps personalized latents to experts.
Source: §1 · p. 1, §4.1 · p. 3Results
Personalized MoEs outperform dense models with similar active parameters while using 20% of total parameters per token, and their latents and experts support recovery of personalized behavior.
Source: §1 · p. 2Takeaways & Limitations
MoE experts encode distinct personalization features through different routing paths, enabling localization and subsequent intervention for controlled personalization.
Source: §7 · pp. 8–9The experiments use English group-level preferences expressed through synthetic system messages, which may differ from noisier, more entangled real user histories.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
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.
1 Introduction
Dense LLMs are difficult to interpret for personalization because polysemantic neurons hinder localization of the components encoding distinct preferences. The paper studies MoE personalization with crosscoder latents to identify personalized features and their expert routing paths.
1 Introduction
- Dense LLMs are hard to interpret for personalization because polysemantic neurons encode multiple unrelated concepts, complicating localization of relevant components.MoEs sparsely route tokens to selected experts, which can encode more distinct task-specific features.Source: §1 · p. 1
- The method adapts MoEs with two-stage parameter-efficient fine-tuning, then trains crosscoders on paired personalized and base activations to identify latents and map them to experts.Figure 1 depicts paired activations collected from personalized and non-personalized MoEs for crosscoder training.Source: §1 · p. 1, Figure 1 · p. 2
- Across three open-weight MoE models and three benchmarks, personalized MoEs outperform dense models with similar active parameters and use only 20% of total parameters per token.They achieve performance competitive with dense models of similar total parameter size.Source: §1 · p. 2
- MoE experts encode distinct personalization features and exhibit lower polysemanticity than dense models, enabling more interpretable feature localization.Crosscoders identify a small set of personalized latents, and causal interventions show these latents suffice to recover personalized behavior.Source: §1 · p. 2
2 Related Work
Prior work personalizes LLMs through user histories, prompts, adapters, reward models, policy merging, and activation steering, while evaluating both group- and individual-preference settings. Related MoE and interpretability research motivates studying how expert routing and crosscoder latents represent personalization.
2.1 Personalization of LLMs
- Personalization methods condition on histories or profiles (Richardson et al., 2023), encode personas in prompts (Jiang et al., 2024b), train adapters, or steer activations.LaMP and LongLaMP evaluate per-user preferences (Kumar et al., 2024), while Multifaceted Collection and Bench use group preferences.Source: §2.1 · p. 2
2.2 Mixture-of-Experts LLMs.
- MoE research studies routing specialization and expert-based adaptation, while interpretability work uses sparse autoencoders, probing, and crosscoders to analyze features and fine-tuned models.This paper compares personalized MoEs with their base models and maps resulting latents onto routed experts.Source: §2.2 · p. 3
3 Preliminaries
The preliminaries formulate personalization as generating a response from a query and user preference, then describe MoE routing and crosscoders as the modeling and analysis tools. The crosscoder jointly reconstructs paired personalized and base-model activations, with model-specific decoder vectors supporting latent identification.
3.1 Problem Formulation
- Given a query and user preference, a personalized LLM generates a response scored against a gold response, with an adapter injecting the preference into generation.The paper realizes the adapter using MoE parameter-efficient fine-tuning and decomposes its encoded information into interpretable latents.Source: §3.1 · p. 3
3.2 Mixture-of-Experts LLMs
- An MoE layer replaces a dense feed-forward block with experts and routes each token to its top-k experts.The routing mechanism is described as selecting experts for each token.Source: §3.2 · p. 3
3.3 Crosscoder
- A crosscoder learns one dictionary of latents that jointly reconstructs paired personalized and non-personalized activations, using a separate decoder vector for each model.Latents whose personalized decoder contributes mainly to reconstructing the personalized model can be identified as personalized features.Source: §3.3 · p. 3
4 Methodology
The method first tunes MoE routers to find experts relevant to personalization, then tunes selected experts. It trains crosscoders on paired activations and maps personalized latents to experts using routing changes and criteria for latent distinctiveness and mapping stability.
4.1 Personalization of MoE LLMs
- The two-stage adaptation first tunes routers with language-modeling loss, then fine-tunes only the most frequently selected experts per layer.This addresses load balancing during pretraining, which spreads tokens evenly across experts and discourages task-specific specialization.Source: §4.1 · pp. 3–4
- Expert selection measures how often tuned routers select experts for tokens, with frequencies normalized so they sum to one at each layer.The method then fine-tunes the k most selected experts per layer.Source: §4.1 · p. 4
4.2 Disentangling Personalized Features
- For disentanglement, the crosscoder is trained on paired personalized and non-personalized activations to identify personalized-specific latents and map them to experts.The analysis selects a middle layer with high routing divergence and low routing entropy.Source: §4.2 · p. 4
- Latents are classified by relative decoder-norm differences, and experts are ranked by their increased selection frequency when a latent fires.The mapping parameters are chosen by grid search for distinctiveness across latents and stability across random subsamples; overly large subsamples can reduce distinctiveness through overlap.Source: §4.2 · p. 4
5 Experimental Settings
The experiments personalize three instruction-tuned MoE models and evaluate them on group- and user-preference benchmarks. Dense models matched for active or total parameter count and fully fine-tuned reference models provide comparisons.
Data and evaluation
- Evaluation covers MULTIFACETED BENCH, LaMP, and LongLaMP, testing group-level preferences and preferences inferred from user interaction histories.MULTIFACETED BENCH has no preferences in common with training; email tasks are omitted because their data is not publicly available, and METEOR and ROUGE-L are reported.Source: §5 · p. 4
- Models are personalized and crosscoders trained on MULTIFACETED COLLECTION, which pairs queries with group-preference system messages and preference-aligned responses.The system messages cover style, background knowledge, informativeness, and harmlessness; responses are generated by GPT-4-Turbo-0125.Source: §5 · p. 4
Models and baselines
- Three instruction-tuned MoE models are evaluated without adaptation, after router tuning, and after router-plus-expert tuning against dense models matched for active or total parameter count.The MoE models are OLMoE-1B-7B, Qwen3-30B-A3B, and Mixtral-8x7B; the dense comparisons are OLMo-2-1B/7B, Qwen2.5-3B/32B, and Llama-2-13B/70B, respectively.Source: §5 · p. 4
- JANUS-7B and its ORPO/DPO variants, fully parameter-fine-tuned on MULTIFACETED COLLECTION, serve as a performance upper-bound reference for similarly sized LLMs.The section also describes Figure 2 as box plots of average performance across personalized and non-personalized variants on seven datasets, with means marked by diamonds.Source: §5 · p. 4, Figure 2 · p. 5
Polysemanticity
- Figure 3 compares k-sparse probe balanced accuracy between MoE and dense LLMs with similar active parameter counts.The caption identifies the comparison but does not report its outcome.Source: Figure 3 · p. 5
6 Results
Personalized MoE LLMs perform competitively against dense models while activating only a fraction of their parameters. The analysis finds lower polysemanticity, interpretable preference latents, distinct expert mappings, and causal evidence that these latents and experts recover aspects of personalized behavior.
6.1 Personalization performance comparison
- Across model families and metrics, MoEs outperform dense LLMs with similar active parameters; Qwen and Mixtral also match or exceed similar-total-parameter dense models.After personalization, MoEs outperform or match dense models with similar active parameters; gains are largest on group-level preferences, and router tuning contributes most in Qwen3 while expert tuning contributes more in Mixtral.Source: §6.1 · p. 5
6.2 RQ1: Do Sparse MoE LLMs offer lower polysemanticity, enabling better disentanglement of personalized features?
- MoE experts outperform similar-active-parameter dense models on balanced-accuracy probes, with the gap remaining wide at k = 1.The authors interpret the k = 1 result as evidence of monosemantic behavior for personalization; same-family comparisons show a wider gap for Olmo and Qwen.Source: §6.2 · pp. 5–6
- Within the same model family, MoE personalized latents generally have lower ν than dense-model latents, except Mixtral versus a similar-active-parameter Llama-13B.The authors interpret lower ν as less contribution to reconstructing the non-personalized model relative to the personalized model; they limit this comparison to the same model family.Source: Figure 4 · p. 5, §6.2 · p. 6
6.3 RQ2: What are the different personalized features disentangled into a set of interpretable concepts?
- The crosscoder identifies 157/65536 Qwen3, 132/131072 Mixtral, and 111/65536 OLMoE personalized latents; most OLMoE and Qwen3 latents fire mainly on system messages.That system-message pattern applies to 85% of OLMoE and 82% of Qwen3 latents; Mixtral is the exception, with 41% firing mainly in responses.Source: §6.3 · pp. 6–7
6.4 RQ3: How are different experts in MoE uniquely aligned with different personalized features?
- For Qwen3, latent–expert mappings achieve stability S = 0.81 and distinctiveness D = 0.94, supporting alignment of different experts with distinct personalization features.The reported mapping uses the optimal 400 examples per latent and restricts routing toward one expert at every layer.Source: §6.4 · p. 7, Figure 8 · p. 7
6.5 Causal analysis
- On MULTIFACETED BENCH, patching personalized latents yields the largest KL-divergence decrease among baselines, supporting their sufficiency for explaining personalized features.Steering dimension-specific top-k′ experts in non-personalized MoEs also yields lower KL divergence than random steering, supporting distinct routing paths.Source: §6.5 · p. 8, Figure 10 · p. 8
7 Discussion and Conclusion
The paper presents a mechanistic study of personalization in MoE LLMs, combining router-and-expert adaptation with crosscoder interpretability. Its findings link MoE personalization to competitive parameter-efficient performance, less polysemantic representations, and distinct routing paths for personalized features.
7 Discussion and Conclusion
- The paper presents, to the authors’ knowledge, the first mechanistic study of personalization in MoE LLMs.It combines two-stage router-and-expert adaptation with crosscoder-based interpretability analysis.Source: §7 · p. 8
- MoE LLMs achieve competitive personalization performance with dense LLMs while using approximately 20% of total parameters.The study also reports lower polysemanticity and improved interpretability of encoded personalization features compared with dense LLMs.Source: §7 · p. 8
- MoE LLMs encode distinct personalization features through different routing paths, enabling localization and subsequent intervention for controlled personalization.Source: §7 · pp. 8–9
Limitations
The study’s conclusions are bounded by its model comparisons, data, and analysis methods. In particular, architectural effects are not fully isolated, and its controlled group-preference setting and interpretability measures have specific limitations.
Models and comparisons
- Architectural effects cannot be fully separated from training differences because dense counterparts differ in pre-training data and recipes.The study analyzes one layer per model, and its conclusions may not transfer to MoE designs with shared experts or finer expert segmentation.Source: Limitations · p. 9
Data and notion of persona
- The experiments cover English group-level preferences represented by synthetic system messages, not individual-user preferences.Real user-history signals may be noisier and more entangled than in this controlled setting.Source: Limitations · p. 9
Crosscoder fidelity
- Crosscoders reconstruct activations imperfectly, leaving reconstruction-error information invisible to latent-level analyses.Variance explained ranges from 0.80 for OLMoE to 0.65 for Mixtral, and personalized-latent selection depends on the decoder-norm threshold.Source: Limitations · p. 9
Probing
- Sparse probing labels individual tokens although persona concepts describe whole responses, introducing label noise and lowering absolute scores.Selecting the best among more MoE components may favor MoE models; probes measure decodability, not whether the model uses a concept.Source: Limitations · p. 9
Expert attribution and causality
- Expert sets derived from routing frequencies at firing tokens are correlational and depend on example count, selection rule, and background routing rate.Source: Limitations · p. 9
Evaluation of generation
- ROUGE-L and METEOR are imperfect proxies for personalization, and sampling-based generation adds variance.Automated explanations also rely on an LLM explainer and scorer whose biases may shape descriptions.Source: Limitations · p. 9
Ethical Considerations
The authors report using Claude for language polishing, grammatical corrections, clarity, and coherence; they state that technical contributions and final decisions were theirs.
Ethical Considerations
- Claude API assisted with language polishing, grammar, clarity, and coherence; the authors made the technical contributions and final decisions.Source: Ethical Considerations · p. 9
A Relevant Model Architecture Details
The crosscoder learns a shared dictionary of interpretable concepts from personalized and base-model activations, reconstructing both from sparse latent activations. The appendix also specifies its sparsity and expansion settings, comparison scope, and generation configuration.
A.1 Crosscoder
- The crosscoder learns a shared dictionary of interpretable concepts that explains and reconstructs activations from both personalized and non-personalized models.Its encoder maps model activations into a higher-dimensional latent space.Source: §A.1 · pp. 11–12
- The crosscoder uses an expansion factor of 32, with latent dimension D = α × d.Here d is the model dimension.Source: §A.1 · p. 11
- Table 1 compares each MoE LLM with dense LLMs of similar active and total parameter sizes, using the same model family when available.For dense models, active and total parameter counts are the same.Source: Table 1 · p. 12
- BatchTopK enforces L0 sparsity by selecting the top-k latents per example across a batch, limiting the batch to nk activated latents.The authors adopt it following prior work on obtaining better-decoupled latents with L0 rather than L1 sparsity.Source: §A.1 · p. 12
B.1.1 Hyperparameters
- MULTIFACETED BENCH generation uses temperature 1.0, top-p 0.9, repetition penalty 1.03, and a 4096-token generation limit.These settings are specified for inference on that dataset.Source: §B.1.1 · p. 12
B.1.2 Implementation Details
The implementation uses LoRA-PEFT for MoE router and expert tuning and for dense-model personalization, and trains crosscoders on paired residual MLP activations. The appendix figures and tables describe routing analyses, intervention comparisons, benchmark metrics, and examples of personalized latents.
B.1.2 Implementation Details
- LoRA-PEFT is used to tune MoE routers and personalized experts and to personalize dense LLMs; only MLP weights are updated.Crosscoders use L0 BatchTopK with expansion factor 32 on paired residual MLP contributions.Source: §B.1.2 · p. 12
- Routing analyses identify L6 as the best layer for OLMoE-1B-7B and L15 for Mixtral-8x7B, with relative-norm distributions shown for latents from those layers.The routing figures plot model-to-model routing JS-divergence and entropy of overlapped experts.Source: Figure 11 · p. 12, Figure 12 · p. 13, Figure 13 · p. 13
- The (N, k) grid searches assess top-k′ expert distinctiveness across latents and resampling stability for a particular latent in OLMoE-1B-7B and Mixtral-8x7B.Each model has a separate grid-search figure.Source: Figure 15 · p. 13, Figure 16 · p. 14
- KL patching compares personalized latents against frequency-matched random and shared latents for reducing next-token KL divergence to the personalized MoE, with significance marked at p < 0.001 versus both controls.The figure reports the first nine response tokens, averages three control draws, and uses 95% bootstrap confidence intervals over 921 MULTIFACETED BENCH examples.Source: Figure 17 · p. 14
- Table 2 reports METEOR and ROUGE across personalization benchmarks, while Table 3 gives latent explanations, preference dimensions, associated experts, routing shares, and example sequences.Table 2 distinguishes active-size and total-size comparisons and identifies best and second-best results within each model family, benchmark, and metric.Source: Table 2 · p. 15, Table 3 · p. 16