User LMs for Personalized Simulation

• User Language Models are specialized systems designed to replicate realistic user linguistic behavior by capturing idiosyncratic phrasing, partial intent, and multi-turn dialogue dynamics.
• They employ memory-augmented prompting and embedding-based personalization techniques that yield statistically significant gains, such as up to 7 points improvement in prediction accuracy.
• User LMs underpin applications in recommendation, simulation, and dialogue evaluation by modeling inter-user differences and enhancing personalized experiences in real-time interactions.

User LLMs (User LMs) are a class of LLMs designed, trained, or adapted specifically to represent, simulate, and respond to human users’ linguistic behavior, preferences, profiles, or intents. Unlike traditional assistant LMs, which are optimized for clear, exhaustive, and helpful responses, User LMs are explicitly optimized for realism in user communication—capturing idiosyncratic phrasing, partial intent revelation, and heterogeneity in subjective behavior. As User LMs underpin critical applications in conversational AI, recommendation, simulation, and personalization, research in this domain has evolved rapidly to answer challenging questions regarding user-specific modeling, multi-domain heterogeneity, evaluation, and cross-user differentiation.

1. Foundations and Motivation

The conceptual foundation for User LLMs stems from the observation that conversational and interactive LLMing fundamentally involves two distinct participants: the user and the assistant. While state-of-the-art LMs are post-trained to respond in a cooperative, polished manner as assistants, the utterances and behaviors of real users are markedly different—often partial, ambiguous, and contextually situated. Empirical studies demonstrate that direct reuse of assistant LMs as user simulators fails, as their helpfulness bias markedly reduces realism in user simulation environments (Naous et al., 8 Oct 2025). This gap motivates the explicit development of User LMs, defined as models post-trained or architected to mirror natural user utterances and behavioral dynamics in multi-turn dialogues and personalized applications.

The evolution of User LMs can be traced to four primary drivers:

• The inability of assistant LMs to realistically simulate user behaviors for system evaluation (Naous et al., 8 Oct 2025).
• The need for evaluating and benchmarking assistants under “real user” conditions, revealing true performance boundaries.
• Advances in user-centric benchmarking and the emergence of data collection protocols capturing real user intents, profiles, and interactions (Wang et al., 22 Apr 2024, Wang et al., 16 Jan 2024).
• The recognition that personalization—whether in recommendations, dialogue, or content generation—demands modeling user-level variance that static group-based or instruction-tuned LMs cannot provide (Hwang et al., 2023, Qiu et al., 4 Mar 2025, Qiu et al., 28 Jul 2025).

2. Principles of User Modeling: Personalization, Memory, and Inter-User Difference

A central paradigm in User LM research is personalization—the adaptation of model outputs to individual user characteristics based on past interaction, dynamic intent, or latent behavioral signals. Several canonical strategies have been proposed:

• Explicit Use of Past Opinions/Responses: A user's historical responses (either generated or approved) serve as the core signal for predicting future user behavior or preference. Empirical analyses confirm that these historical responses (when appended or retrieved as top‑k most relevant examples) yield statistically significant gains (up to 7 points in prediction accuracy) over demographic or ideological group-based prompting alone (Hwang et al., 2023). This memory-based personalization is formalized as $p(y|x, \mathcal{P}_u; \theta)$, where $\mathcal{P}_u$ denotes the user’s profile of past utterances or responses (Wu et al., 22 Jun 2024).
• Profile Construction and Compact Embeddings: User histories, whether represented as free-form text, ratings, or behavioral logs, can be compressed using encoder architectures (transformers, UEMs, or autoencoders) into fixed-length dense vectors serving as soft prompts or keys in attention mechanisms (Doddapaneni et al., 10 Jan 2024, Ning et al., 21 Feb 2024, Qiu et al., 28 Jul 2025). These embeddings encapsulate user-specific patterns while enabling efficient integration with LLMs at inference.
• Inter-User Difference Modeling: Despite gains from modeling individual history, research demonstrates that leveraging systematic inter-user comparison is critical for true personalization. Approaches like Difference-aware Personalization Learning (DPL) (Qiu et al., 4 Mar 2025) and Difference-aware Embedding-based Personalization (DEP) (Qiu et al., 28 Jul 2025) extract, in task-aware fashion, the dimensions along which a user’s responses systematically diverge from peer users engaging with the same content. This is operationalized either via prompt-based structured difference extraction or by constructing latent-space soft prompts through embedding contrasts and sparse autoencoding.
• Clustering and heterogeneity-aware training: For domains with inherently subjective or highly heterogeneous user behaviors (e.g., idiosyncratic browsing “languages”), clusterwise modeling (HeTLM) assigns users to clusters and trains predictor heads per cluster, reducing within-group variance and improving personalization mean performance (Sundaresan et al., 21 Aug 2025).

3. Model Architectures and Personalization Mechanisms

The architectural design of a User LM incorporates several mechanisms geared toward efficient and expressive user representation:

• Memory-Augmented Prompting: Relevant past opinions or behavioral snippets are retrieved and concatenated with current prompts, often following a retrieval-based or similarity-driven top‑k selection to avoid prompt noise (Hwang et al., 2023, Wu et al., 22 Jun 2024).
• Embedding-Based Integration: Transformer-based User Embedding Modules (UEMs) compress long user histories into compact vectors, which serve as soft prompts prepended to, or integrated via cross-attention with, the standard LM input embeddings (Doddapaneni et al., 10 Jan 2024, Ning et al., 21 Feb 2024). The cross-attention is typically mathematically described as $Attention(Q, K, V) = Softmax((QK^\top)/\sqrt{d})V$, where user embeddings serve as keys and values.
• Latent Difference Encoding: DEP, for example, computes a user’s behavioral difference with respect to peer groups in the latent embedding space and projects these signals into the LM’s input layer via a sparse autoencoder and lightweight projection network. This composite prompt informs the LM of both typical and distinct user behavioral patterns (Qiu et al., 28 Jul 2025).
• Clustered Parameterization: Heterogeneity-aware models such as HeTLM operate multiple predictor LMs over endogenously induced user clusters and update cluster assignments online, improving both average personalization and reducing population-wide variance (Sundaresan et al., 21 Aug 2025). In large-scale setups, fine-tuning SLMs (Small LLMs) with persona-specific low-rank adapters ensures both computational tractability and user-level calibration (Thakur et al., 18 Aug 2025).

4. Evaluation Methodologies and Benchmarks

Robust evaluation of User LMs leverages both intrinsic measures—assessing similarity to human user utterances—and extrinsic tests that quantify the downstream impact on system performance:

• Distributional Alignment and Perplexity: Metrics such as token-level PPL on human user utterances confirm that User LMs trained from base LMs (not instruction-tuned) match empirical user distribution more accurately than assistant-based simulators (Naous et al., 8 Oct 2025).
• Multi-Turn Diversity and Intent Revelation: Evaluations measure 1-gram diversity in opening utterances, intent decomposition (cumulative n-gram overlap with the high-level user intent), and intent coverage (fraction of atomic intent-relevant information revealed during conversation).
• Role and Intent Adherence: Automated routines assess whether the model consistently occupies the user’s role and maintains original intent throughout conversation (Naous et al., 8 Oct 2025).
• Dialogue Termination: Precision, recall, and $F_1$ score for conversation-ending cues determine the naturalness of termination; User LMs outperform assistant LMs which rarely terminate conversations promptly.
• Personalization Metrics: ROUGE, METEOR, BERTScore, and unique LLM-based metrics (e.g., S-72B, S-GPT) are used to benchmark personalized review generation (Qiu et al., 4 Mar 2025, Qiu et al., 28 Jul 2025). Additionally, contrastive loss and domain importance weighting are used in multi-domain recommendation contexts (Bao et al., 7 Jul 2025).
• User-Centric Benchmarks: Datasets and scoring protocols aligned with real user intents and satisfaction (e.g., URS (Wang et al., 22 Apr 2024), CLUE (Liu et al., 21 Feb 2025), user survey-based studies (Wang et al., 16 Jan 2024)) quantify both subjective and objective experience with User LMs and their downstream assistants.
• Simulation Robustness and Task-Downstream Impact: Simulating coding, math, and other multi-turn interactions with User LMs reveals substantial performance drops in even strong assistants when exposed to realistic user utterances, quantifying the value of accurate user simulation for rigorous system evaluation (Naous et al., 8 Oct 2025).

5. Applications: Recommendation, Simulation, Dialogue, and Beyond

User LMs are applied across diverse settings where modeling individual user behavior is essential:

• Personalized Recommendation Systems: Multi-modal embeddings, clustering, and difference-aware textual representations yield notable gains in next-item recommendation, user intent prediction, and cold-start scenarios (Doddapaneni et al., 10 Jan 2024, Xu et al., 18 Jan 2025, Zhou et al., 23 Feb 2024, Bao et al., 7 Jul 2025).
• Conversational Simulation: Purpose-built User LMs reveal assistant model weaknesses, provide harder evaluation benchmarks, and drive improvements in robust assistant design (Naous et al., 8 Oct 2025). Synthetic user simulation using LLMs further enables scalable, attribute-diverse tests in dialogue systems (Ahmad et al., 18 Feb 2025).
• Profiling and Real-Time Personalization: Techniques for constructing and dynamically updating structured user profiles from free-text or biographical data amplify personalization in large-scale systems (Prottasha et al., 15 Feb 2025). Efficient soft prompt construction (via UEMs or autoencoders) supports adaptation over long user histories (Ning et al., 21 Feb 2024, Doddapaneni et al., 10 Jan 2024).
• User Experience Analysis: LLM-powered interviewing and user intent modeling (CLUE, taxonomy-driven logs) surface nuanced user opinions, model satisfaction, and drive user-centric evaluation and model design (Liu et al., 21 Feb 2025, Wang et al., 16 Jan 2024). Additional studies address alignment with diverse user demographics and linguistically diverse populations (Knoeferle et al., 18 Feb 2025, Panda, 23 Sep 2024).

6. Technical and Methodological Challenges

Development and deployment of User LMs expose several open challenges:

• Idiosyncratic and Heterogeneous Preferences: Modeling individual users’ highly subjective behaviors, particularly across open domains or in browsing scenarios, requires explicit mechanisms for knowledge compression (e.g., customized tokens, masking), robust cross-domain knowledge fusion, and clustering to address population variance (Sundaresan et al., 21 Aug 2025, Bao et al., 7 Jul 2025).
• Scalability and Efficiency: Methods such as sparse autoencoders, low-rank adapters, and embedding compression facilitate scalable training and inference, allowing fine-grained personalization at population scale (Qiu et al., 28 Jul 2025, Thakur et al., 18 Aug 2025).
• Data Sparsity and Cold-Start: Language-based user profiling and textual distillation from frozen LLMs deliver interpretable summaries and address cold-start performance issues without reliance on high-dimensional latent vectors (Zhou et al., 23 Feb 2024, Thakur et al., 18 Aug 2025).
• Bias, Fairness, and Echo Chambers: Over-reliance on past opinions raises the risk of bias reinforcement; hybrid strategies—e.g., simulated annealing that shifts between group-level and individual-level alignment—are proposed to balance personalization with diversity (Hwang et al., 2023).
• Position and Tokenization Effects: Performance is sensitive to both the order and type of user profile elements included in the context. Placing more relevant personalized responses at the beginning of the prompt, and using specialized tokenizers for browsing “languages”, strengthens alignment (Wu et al., 22 Jun 2024, Sundaresan et al., 21 Aug 2025).
• Evaluation Complexity: Realistic evaluation necessitates both automated statistical measures and human-centered feedback, leveraging diverse metrics and intent-aligned benchmarks for comprehensive assessment (Wang et al., 22 Apr 2024, Liu et al., 21 Feb 2025, Wang et al., 16 Jan 2024).

7. Future Directions

The User LM field is undergoing rapid development in several directions:

• Unified and Multimodal Representations: There is interest in extending current frameworks to encode multimodal user signals beyond text, including images, audio, or behavioral traces (Doddapaneni et al., 10 Jan 2024, Xu et al., 18 Jan 2025).
• Parameter-Efficient Personalization: Techniques such as LoRA, prompt-tuning, and efficient autoencoders enable dynamic personalization with minimal resource overhead (Thakur et al., 18 Aug 2025, Doddapaneni et al., 10 Jan 2024).
• Rich Cross-User Difference Modeling: Advancements in latent space contrastive mechanisms and heterogeneous domain modeling continue to drive performance on highly personalized and open-domain tasks (Qiu et al., 4 Mar 2025, Bao et al., 7 Jul 2025, Qiu et al., 28 Jul 2025).
• User-Guided Adaptivity and Steerability: Empowering users to inspect, adjust, and steer profile summaries or personalize interaction style holds both practical and ethical value, facilitating trust and system transparency (Zhou et al., 23 Feb 2024).
• Enhanced Simulation Environments: Next-generation User LMs will set more realistic, challenging standards for assistant model development, closing the gap between offline benchmarks and real-world user experience (Naous et al., 8 Oct 2025).

The orchestration of these technical principles is shaping the future architecture, evaluation, and deployment of language technology systems that are both robust to human heterogeneity and optimized for real-world interaction fidelity.