persona design

This work addresses the limitation of existing personalization methods for large language models, which rely on explicit preference annotations and thus struggle in natural interactive settings. We propose the first dynamic user profiling framework that leverages implicit behavioral signals without requiring explicit feedback. Our approach employs a closed-loop, behavior-prediction-driven iterative mechanism, integrating memory augmentation with dynamic representation learning to continuously model user preferences. To evaluate performance comprehensively, we introduce a novel protocol encompassing behavior prediction accuracy, profile stability, and decision-making fidelity. Evaluated on real-world Reddit data, our method achieves a 61.0% decision prediction accuracy, significantly outperforming static profiling, pure memory-based retrieval, and non-personalized baselines.

This study systematically investigates the persistent performance degradation of role-playing large language models (LLMs) in ultra-long dialogues (>100 turns), focusing on dynamic decay across three dimensions: role fidelity, instruction adherence, and safety. We introduce the first dialogue-conditioned long-horizon evaluation protocol, benchmarking seven prominent open- and closed-source models using long-context modeling and multi-dimensional dynamic quantitative metrics. Our analysis reveals, for the first time, a fundamental long-term trade-off between role fidelity and instruction adherence: all models exhibit significant erosion of role consistency as dialogue length increases—particularly in goal-directed scenarios—where responses progressively converge toward role-agnostic baselines, confirming a structural failure in long-term role persistence. These findings expose an intrinsic fragility in current role-playing paradigms and establish a reproducible benchmark with actionable insights for developing trustworthy, long-interaction role-aware LLMs.

To address the challenges of sustaining persona model optimization and limited behavioral prediction accuracy in dynamic real-world scenarios, this paper proposes an iterative reinforcement learning (RL)-based targeted refinement framework. Methodologically, it introduces a novel direction-search-enhanced RL paradigm that automatically identifies effective update pathways by modeling behavior-prediction discrepancies, enabling interpretable, bias-driven persona evolution. Furthermore, it integrates large language models (LLMs) to construct fine-grained, temporally aware persona representations and adaptive update mechanisms. Experimental evaluation across 10 domains involving 4,800 users demonstrates that, after four iterative refinement rounds, the framework reduces average behavioral prediction error by 32.2%, outperforming the best baseline by 22.92%. This work advances persona modeling through explainable, discrepancy-guided adaptation and LLM-powered temporal representation learning.

Existing LLM-based user profiling primarily relies on demographic attributes or historical behavioral data, failing to model the underlying cognitive motivations behind user judgments—thus limiting preference prediction accuracy. To address this, we propose a psychology-enhanced LLM persona construction framework that, for the first time, integrates computable Big Five personality traits and the Original World Beliefs theory into a structured psychological scaffold, explicitly modeling user reasoning via a rational generation mechanism. Our approach synergizes theory-of-mind modeling, structured prompt engineering, and multi-source rationality fusion. Empirical evaluation on public opinion and movie preference prediction tasks demonstrates significant improvements over state-of-the-art baselines. Moreover, the generated rational explanations achieve cognitive plausibility comparable to human-authored ones. This work validates the effectiveness and novelty of theory-driven modeling in advancing LLMs’ capacity for deep user understanding.

To address challenges in personalized response prediction for large language models—including difficulty modeling user preferences, low retrieval efficiency under resource constraints, and poor cold-start performance—this paper proposes a hierarchical collaborative data representation framework. Methodologically, it designs a hierarchical vector index and database structure to jointly encode general knowledge and individual preferences; introduces user collaborative representation learning and context-aware embedding refinement to mitigate data sparsity via cross-user knowledge complementarity; and integrates retrieval-augmented generation (RAG) for efficient low-rank retrieval. Its primary contribution is the first collaborative data refinement paradigm, significantly enhancing cross-task generalization and knowledge sharing. Experiments demonstrate stable response accuracy even with minimal retrieval size, over 10% improvement in cold-start scenarios, and substantially reduced context-length requirements—yielding exceptional practicality in long-history and limited-context settings.

Personalized prompting offers large opportunities for deploying large language models (LLMs) to diverse users, yet existing prompt optimization methods primarily focus on task-level optimization while largely overlooking user-specific preferences and latent constraints of individual users. This gap is primarily due to (i) the absence of high-quality, privacy-sensitive data that capture personalized user-LLM interactions at scale, and (ii) the lack of robust reward signals for individual preferences. To overcome existing data limitations, we introduce a high-fidelity synthetic data generation framework called PersonaGym. Unlike prior work that treats personalization as static persona-preference pairs, PersonaGym models a dynamic preference process via an agentic LLM system to simulate realistic preference behaviors and semantic-aware noise in order to generate personalized multi-turn interaction trajectories. Using PersonaGym, we release PersonaAtlas, a large-scale, high-quality, and diverse synthetic dataset of high-fidelity multi-turn personalized interaction trajectories that closely mirror real-world preference expression and noise patterns. We further propose Personalized Prompt Optimization (PPOpt), a scalable and model-agnostic framework that optimizes user prompts based on interaction histories without modifying the deployed LLM. PPOpt adopts a reason-then-optimize paradigm that infers an explicit user profile and conditions prompt rewriting on the user profile to avoid reward hacking. Our training procedure for PPOpt integrates a cold-start supervised prior with outcome-driven multi-objective reinforcement learning. We present extensive experiments to demonstrate consistent improvements over state-of-the-art baselines in terms of task performance, personalization quality, and robustness to noisy as well as to sparse preference signals.

Current LLM-based user simulators often exhibit overly cooperative and homogeneous behaviors, failing to capture the diversity and challenge inherent in real users, which undermines agent performance in practical settings. This work proposes Persona Policies (PPol), a novel framework that models user personas as evolvable Python programs. By integrating multi-objective evolutionary algorithms with LLM-guided program synthesis, PPol automatically generates task-consistent, highly realistic, and behaviorally diverse user simulators without manual design. Evaluated on the tau²-bench benchmark, PPol improves simulator realism by 33–62% and achieves an 80.4% human misclassification rate, indicating high behavioral fidelity. Agents trained with PPol demonstrate a 17% relative improvement in task success rates, highlighting its effectiveness in enhancing agent robustness through more authentic user interactions.

Traditional user profiling approaches rely on discriminative models and manual feature engineering, struggling to capture long-tail behaviors and often yielding fragmented, logically inconsistent profiles. This work proposes UserGPT, a novel framework that leverages large language models to transform massive, noisy user behavioral logs into coherent user narratives, enabling holistic personality inference. UserGPT introduces a dual-path paradigm—combining attribute generation and summary generation—and integrates a user behavior simulation engine, a data semanticization module, multi-stage supervised fine-tuning, and a dual-filter grouped relative policy optimization (DF-GRPO) strategy. Evaluated on HPR-Bench, the framework achieves an Avg@10 of 0.7325 for label prediction and an Acc_Ex of 0.7528 for summary generation, while compressing behavioral records by 97.9% without significant loss of critical information.

Existing agent evaluation benchmarks struggle to assess the ability to model user preferences and proactively acquire information during long-term, fragmented interactions. To address this gap, this work introduces the first benchmark specifically designed for evaluating agents in sustained user collaboration. It systematically incorporates two key evaluation dimensions—personalization modeling and proactive behavior—through temporally structured, heterogeneous task sequences. The benchmark features interactive tasks embedded with dynamic user preferences and an extensible memory interface, enabling controlled comparisons across diverse memory architectures. Evaluations with state-of-the-art large language models reveal limited performance in authentic personalized scenarios, highlighting significant bottlenecks in continuously modeling evolving preferences and proactively completing missing information, thereby offering clear directions for future research.

This work addresses the challenges of intent entanglement and noise in user behavior logs by proposing a hierarchical multi-role profiling framework. The approach first aggregates behavioral logs to construct an intent memory, then employs hierarchical clustering coupled with natural language labeling to generate multiple evidence-grounded roles. Role generation is formulated as an optimization problem, and the framework introduces a novel group-based Direct Preference Optimization (Group-based DPO) training strategy to enhance role consistency, alignment with supporting evidence, and overall authenticity. Experimental results on large-scale service logs and two public datasets demonstrate that the generated roles significantly outperform baseline methods in coherence, credibility, and evidential grounding, while also yielding improved performance in future interaction prediction tasks.