UserGPT Technical Report

Abstract: Personalized user understanding from large-scale digital traces remains a fundamental challenge. Traditional user profiling methods rely on discriminative models and manual feature engineering to predict discrete attributes, often producing fragmented and logically inconsistent profiles that generalize poorly to long-tail behaviors. In this work, we study a generative paradigm in which LLMs summarize long and noisy behavioral histories into coherent narratives that capture nuanced user evolution. Our experiments show that even strong LLMs remain limited in complex and implicit personalization reasoning. We propose UserGPT, a framework for improving LLM-based persona understanding through both attribute generation and summary generation. To address the scarcity of real-world behavioral data, we develop a User Behavior Simulation Engine that produces realistic and complex user trajectories. We further introduce a Data-Centric Semantization module that transforms heterogeneous behavioral logs into structured and semantically coherent inputs, reducing noise and sparsity. On top of this pipeline, we design a curriculum-driven post-training strategy that combines multi-stage Supervised Fine-Tuning (SFT) with Dual-Filter Group Relative Policy Optimization (DF-GRPO) to strengthen reasoning over long behavioral histories. We also construct HPR-Bench, a benchmark for holistic persona reasoning derived from simulated data. On HPR-Bench, UserGPT achieves an Avg@10 score of 0.7325 on tag prediction and an $Acc_{Ex}$ score of 0.7528 on summary generation, while compressing behavioral records by up to 97.9% with critical information preserved. These results demonstrate the effectiveness of UserGPT for holistic persona reasoning and personalized user-agent interaction.

• The paper introduces UserGPT, showcasing an end-to-end framework that employs generative user profiling through curriculum-driven adaptation and reinforcement learning for enhanced logical consistency.
• It leverages a user behavior simulation engine and multi-source data refinement to achieve significant token reduction and accurate profile summarization across diverse behavioral inputs.
• Empirical evaluations on HPR-Bench demonstrate improved atomic tag prediction and robust out-of-domain performance, supporting scalable, real-time personalization.

Comprehensive Technical Review of "UserGPT Technical Report" (2605.08766)

Motivation and Context

Conventional user profiling approaches primarily use discriminative models based on tag-centric prediction, resulting in fragmented and often logically inconsistent profiles that lack adaptability, especially across long-tail user behaviors. As digital interactions proliferate, the limitations of isolated tag classification—in terms of scalability, generalization, and logical consistency—are increasingly evident. Recent advances in LLMs have propelled efforts to extract semantically rich user profiles via generative modeling paradigms, yet naïvely applying LLMs to protracted behavioral histories has proven insufficient for nuanced and temporally consistent persona modeling.

Framework Overview

UserGPT proposes a principled, end-to-end system combining generative user profiling with advanced curriculum-driven adaptation to tackle cognitive and engineering bottlenecks in user understanding. The foundational pipeline comprises:

• User Behavior Simulation Engine: Crafted via persona-driven, environment-interactive, and evolutionary agents inspired by high-dimensional preference modeling (AlignX, SocioVerse), this module generates lifelike, multi-year behavioral trajectories based on real demographic distributions, and injects operational noise for authenticity.
• Data-Centric Semantization: Raw, heterogeneous behavior logs are refined using micro-level entity denoising and macro-level hierarchical structuring. Semantic Refiner leverages a lightweight LLM for entity rewriting and enrichment, systematically stripping promotional noise and augmenting sparsity with external metadata. The resultant Multi-source User Behavior (MUB) corpus achieves >75% reduction in input tokens, optimizing signal-to-noise ratio and model context efficiency.

Curriculum-Driven Post-Training

The modeling layer integrates multi-stage Supervised Fine-Tuning and Reinforcement Learning:

• Multi-stage SFT: Three stages sequentially build atomic attribute inference (explicit behaviors), robustness on controversial/ambiguous cases (implicit cues), and composite profile summarization. Dual verification mechanisms ensure data fidelity and alignment with legacy heuristic labels.
• Reinforcement Learning via DF-GRPO: Dual-Filter Group Relative Policy Optimization incorporates hierarchical filtering (sample/group) to maximize training signal quality and stability, balancing atomic attribute accuracy and summary generation quality without excessive computational escalation.

Prompt engineering is critical throughout, utilizing chain-of-thought (CoT) schema to explicitly require concise, multi-step reasoning and minimize hallucination.

Benchmarking with HPR-Bench

Evaluation is conducted via Holistic Persona Reasoning Bench (HPR-Bench), constructed through rigorous stratification, automated QC, and human expert validation. It measures atomic tag prediction and composite profile summarization across diversified behavioral, demographic, and temporal axes. The benchmark is notable for its coverage, difficulty, and ground-truth reliability.

Empirical Results

UserGPT exhibits strong, quantifiable advances:

• Atomic Tag Prediction: Achieves Avg@10 of 0.7325 on HPR-Benchtag, nearly matching Qwen3.6-Plus, despite operating on a significantly smaller parameter scale (8B vs. >235B+).
• Profile Summarization: Obtains AccEx of 0.7528 (vs. 0.7014 for Qwen3-235B-A22B-Thinking-2507), with information coverage (COVEx) of 0.9747 and up to 97.9% data compression ratio, validating efficient retention of salient personal details.
• Robustness: On out-of-domain tasks, only marginal drops are observed compared to its backbone model, confirming preservation of general capabilities despite domain-specific specialization.
• Ablation Studies: Training exclusively on controversial (hard) samples produces measurable improvements in Pass@1, Pass@5, and Pass@10 scores, emphasizing the necessity of high-value, challenging data. Multi-stage curriculum learning synergistically enhances both atomic and composite reasoning abilities. DF-GRPO filtering is shown to be essential for maximizing RL gains.

Notably, even the largest SOTA LLMs fail to robustly infer complex, implicit, and temporally consistent user attributes, with persistent logical conflicts and incomplete coverage. UserGPT’s workflow resolves many deficiencies via targeted data synthesis and progressive curriculum adaptation.

Incremental Profiling Paradigm

The static profile recomputation paradigm is computationally prohibitive. The report outlines an incremental profiling scheme, where the compressed summary ("pluggable memory") is updated through assimilative reasoning over recent behavioral increments, triggered by both event-driven and periodic signals. This facilitates scalable, real-time personalization and resolves the temporal conflict between long-term stability and short-term intent.

Theoretical and Practical Implications

UserGPT establishes a methodological shift from fragmented tag-based systems toward unified, narrative-driven persona modeling, offering clear advantages in logical consistency, temporal reasoning, and adaptation to evolving user contexts. Its generative summaries enable integration as memory modules for AI agents, supporting more sophisticated, context-aware interactions and bridging the "cold start" gap in application onboarding.

From a theoretical perspective, the approach validates curriculum-driven adaptation and chain-of-thought prompt design as crucial for domain-specific reasoning in LLMs. It also demonstrates the necessity of engineered synthetic data and multi-stage benchmarking for evaluating and advancing user profiling models.

In practice, the incremental profiling module is poised for industrial scalability. The report’s findings spotlight the persistent gaps in LLMs' implicit reasoning, information grounding, and context utilization, framing future research toward multimodal fusion, deeper temporal inference, and hallucination suppression.

Conclusion

UserGPT delivers an authoritative framework for generative, lifelong user profiling, outperforming baseline and SOTA models in complex persona reasoning and efficient information compression. Its integrated pipeline—from synthetic data generation to curriculum-driven post-training and RL optimization—constitutes a comprehensive approach to overcoming both data and modeling bottlenecks. By establishing HPR-Bench as a rigorous evaluation benchmark, the work provides reproducible metrics and standards for advancing LLM-based user profiling.

Looking ahead, further developments should focus on multi-hop inference from sparse events, advanced hallucination mitigation, utilization of extended context windows, and multimodal integration for richer user modeling. This trajectory enables next-generation personalized AI systems with robust, logically consistent, and dynamically evolving user-agent memories.