UserGPT: Generative Persona Reasoning

• UserGPT is a generative framework that reframes profiling as holistic persona reasoning over long, noisy, multi-source behavioral histories.
• It integrates synthetic data generation, semantic compression, curriculum-driven post-training, and reinforcement learning to generate structured attributes and coherent summaries.
• UserGPT overcomes traditional tag-based profiling limitations by enhancing long-tail generalization, reducing fragmented profiles, and supporting open-ended user attributes.

Searching arXiv for the primary paper and closely related work on user behavior foundation models. UserGPT is a framework for personalized user understanding that reframes user profiling as holistic persona reasoning over long, noisy, multi-source behavioral histories rather than as a collection of disconnected tag-prediction tasks. In the technical report "UserGPT Technical Report" (Xuan et al., 9 May 2026), the central claim is that traditional discriminative profiling produces fragmented profiles, logical contradictions, weak long-tail generalization, and poor support for open-ended attributes, whereas a generative LLM can distill behavioral history into both structured atomic attributes and a coherent natural-language summary. The framework combines synthetic data generation, semantically structured behavioral compression, curriculum-driven post-training, and reinforcement learning to support both attribute generation and summary generation. The broader research context includes large-scale behavior foundation modeling, exemplified by BehaveGPT (Gong et al., 23 May 2025), but UserGPT is specifically oriented toward persona understanding and personalized user-agent interaction rather than next behavior prediction.

1. Conceptual framing and problem definition

UserGPT is presented as a shift from tag-based user profiling to a generative paradigm in which an LLM summarizes long and heterogeneous behavioral traces into a coherent, evolving persona memory (Xuan et al., 9 May 2026). The report argues that classical profiling systems are limited by siloed modeling, because each attribute is predicted independently; by logical conflicts, because independent tags can contradict one another; and by weak handling of open-domain traits, because interests, lifestyle, and evolving intent are not naturally reducible to rigid labels.

Within this formulation, the target is not merely to infer isolated properties, but to jointly generate atomic attributes and a summary: $(\hat{y}^a, \hat{y}^s)=f(U)$ where $\hat{y}^a$ denotes atomic attributes and $\hat{y}^s$ denotes the summary (Xuan et al., 9 May 2026). The behavioral corpus $U$ is interpreted as evidence to be synthesized rather than as a feature table to be classified. This suggests a closer alignment with narrative persona construction than with conventional attribute classification.

The report further emphasizes that even strong LLMs remain limited in complex and implicit personalization reasoning. Accordingly, UserGPT is not introduced as a prompt-only system, but as a full pipeline that includes data simulation, semantization, staged supervision, and RL-based alignment (Xuan et al., 9 May 2026).

2. System architecture and end-to-end workflow

The UserGPT workflow is organized into three major layers: data foundation, model post-training, and evaluation (Xuan et al., 9 May 2026). The data foundation contains the User Behavior Simulation Engine, which generates realistic synthetic multi-year logs, and Data-Centric Semantization, which converts heterogeneous raw behavior records into semantically coherent and compact model inputs. The post-training stage applies a curriculum-driven strategy that combines multi-stage supervised fine-tuning with reinforcement learning. Evaluation is conducted on HPR-Bench, a benchmark derived from simulated data for both atomic tag inference and summary generation.

The central architectural idea is that persona reasoning depends jointly on a strong data substrate and a learning curriculum that exposes the model progressively to short, medium, and long behavioral evidence (Xuan et al., 9 May 2026). This progression is explicit: Stage 1 targets basic atomic extraction, Stage 2 targets controversial and boundary cases, and Stage 3 targets composite profile summary generation. The overall design therefore treats user understanding as a reasoning problem over temporally extended traces rather than as a static prediction problem.

A useful comparison point is BehaveGPT, which also argues that user behavior has enough structure and scale to support a foundation-model approach (Gong et al., 23 May 2025). However, BehaveGPT formalizes the task primarily as next behavior prediction, with events represented as

$x_i = (d_i, t_i, l_i, e_i)$

and a prediction target

$$\hat{b}_{t} = f(x_{t-I}, x_{t-I+1}, \ldots, x_{t-1})$$

(Gong et al., 23 May 2025). UserGPT instead targets persona-level inference and summary generation over long, noisy, multi-platform histories (Xuan et al., 9 May 2026).

3. Synthetic data generation and persona evolution

Because real long-term, multi-platform user histories are described as scarce and privacy-sensitive, UserGPT constructs a User Behavior Simulation Engine to produce synthetic user trajectories with realistic complexity (Xuan et al., 9 May 2026). The simulator consists of four modules: Persona-Driven Agent, Environment & Interaction Agent, Simulation Engine with Persona Evolution, and Quality Assurance Mechanism.

The Persona-Driven Agent uses a three-level cognition model comprising Core Persona, Latent Needs, and Specific Intent (Xuan et al., 9 May 2026). Core Persona includes demographics, Big Five traits, and consumption-related psychological needs. Latent Needs are initialized from a desire pool and customized for e-commerce through a “Consumption Desire Library” mined from real reviews. Specific Intent represents context-triggered actions, such as behaviors associated with shopping festivals or social hotspots. The simulator samples a plausible $$(\text{Core Persona}, \text{Latent Need}, \text{Specific Intent})$$ triplet, fills missing details consistently, and advances the synthetic user through stages such as school, graduation, and employment.

The Environment & Interaction Agent creates a dynamic world containing temporal events such as 618, Double 11, back-to-school season, and Spring Festival, as well as spatial trajectories and a cross-platform tool graph spanning e-commerce, delivery, OTA, and POI visits (Xuan et al., 9 May 2026). Persona is explicitly non-static. It evolves through event-driven updates, such as pregnancy, relocation, or job change, and through behavior-driven periodic refinement based on recent activity. This supports lifelong profiling rather than single-snapshot inference.

The simulation engine also contains a Quality Assurance Mechanism based on logical consistency validation, authenticity via noise injection such as misclicks and shared-account scenarios, and manual sample validation (Xuan et al., 9 May 2026). This suggests that the synthetic data is intended not merely to scale the corpus, but to approximate realistic ambiguity and contamination patterns that challenge persona inference.

4. Data-Centric Semantization and behavioral compression

UserGPT treats raw behavioral logs as too noisy, heterogeneous, and lengthy for direct LLM consumption, and addresses this with a semantization pipeline operating at both micro and macro levels (Xuan et al., 9 May 2026). At the micro level, Entity Refinement targets two specific problems: semantic noise and information sparsity. Product titles may include misleading soft labels such as “for pregnant women,” while new entities may lack sufficient metadata. To address this, the system uses a lightweight fine-tuned LLM refiner, Qwen3-1.7B, for entity rewriting and enrichment. Rewriting removes promotional language and preserves canonical identity plus up to three salient attributes; enrichment supplements missing metadata from internal knowledge bases and normalizes entities into a canonical schema (Xuan et al., 9 May 2026).

The report states that this refinement reduces average token length by over 50% (Xuan et al., 9 May 2026). At the macro level, the system constructs a Multi-source User Behavior (MUB) corpus through three stages: multi-stage entity filtering, hierarchical temporal and semantic aggregation, and salience-based summarization and compression. Filtering removes low-value entities using NER, category constraints, text complexity, and stopword lists. Aggregation is both semantic and temporal: long-term history is aggregated monthly or yearly with an emphasis on high-impact actions, while short-term history is kept at daily granularity with richer behavioral detail. Salience-based summarization maps behaviors to a predefined taxonomy, computes frequency as salience, retains Top-K salient categories, and filters synthetic noise and anomalous loops (Xuan et al., 9 May 2026).

The paper states that the overall data reduction rate is over 75% (Xuan et al., 9 May 2026). Later evaluation reports that UserGPT can compress histories by up to 97.9%, with an average summary of about 1.2K tokens from roughly 15K input tokens, and average compression described as about 80–88.5% depending on the summary of the figure (Xuan et al., 9 May 2026). The preservation claim is operationalized through summary-side evaluation, notably COVEx = 0.9747 on HPR-Benchsum for the full model (Xuan et al., 9 May 2026).

5. Curriculum-driven post-training and DF-GRPO

The post-training strategy is explicitly curriculum-driven, beginning with heuristic pseudo-labeling: $\tilde{y}^{a} = RULE(U)$ where $U$ is the user behavior corpus (Xuan et al., 9 May 2026). The report states that the initial set $Q_{init}$ contains 1 million questions, and that samples are filtered out if an LLM already perfectly matches the heuristic label: $\hat{y}^a$0 leaving 490,000 harder questions (Xuan et al., 9 May 2026).

Stage 1 SFT is designed for basic atomic extraction. Five candidate outputs are generated per question, and an atomic extraction operator $\hat{y}^a$1 recovers the predicted atomic label: $\hat{y}^a$2 A sample is kept only if at least four of five outputs agree with the heuristic pseudo-label: $\hat{y}^a$3 This yields roughly 130k high-quality samples (Xuan et al., 9 May 2026).

Stage 2 SFT targets controversial and boundary cases: $\hat{y}^a$4 These samples are resynthesized with a stronger model and carefully verified, producing about 170k samples (Xuan et al., 9 May 2026).

Stage 3 SFT extends the input context from 16K to 36K tokens and trains the model to produce holistic summaries. Around 130k seed users with consistent heuristics are selected, and Stage 2 outputs are used as intermediate references. The retained data is defined as

$\hat{y}^a$5

(Xuan et al., 9 May 2026). According to the report, Stages 1 and 2 teach the model what to infer, while Stage 3 teaches it how to compose those facts into coherent long-form summaries.

After SFT, UserGPT is aligned with DF-GRPO, a variant of Group Relative Policy Optimization. The paper gives the GRPO objective as: $\hat{y}^a$6 with group-normalized advantage

$\hat{y}^a$7

(Xuan et al., 9 May 2026).

The reward is decomposed into format reward, atomic accuracy reward, and summary quality reward (Xuan et al., 9 May 2026). DF-GRPO introduces two filtering layers: sample-level filtering, which removes truncated outputs and malformed generations, and group-level filtering, which discards groups with average reward below $\hat{y}^a$8 or above $\hat{y}^a$9. The report presents this dual-filter design as important for improving summary quality without destabilizing training (Xuan et al., 9 May 2026).

6. Benchmark design, tasks, and empirical results

UserGPT is evaluated on HPR-Bench, or Holistic Persona Reasoning Bench, which is built from simulated data through a five-stage process: data sourcing and sampling, preprocessing and pre-annotation, difficulty stratification, automated quality control, and human verification by at least five experts with majority-vote resolution (Xuan et al., 9 May 2026). The benchmark has two components: HPR-Benchtag for atomic portrait tag inference and HPR-Benchsum for profile summarization.

For tag evaluation, the paper uses Avg@k or Pass@k over 18 atomic tags, with $\hat{y}^s$0 in the main report (Xuan et al., 9 May 2026). For summaries, the metrics are AccEx, COVEx, BLEU-2 / BLEU-4, and Scoresim, along with LLM-as-judge dimensions of completeness, consistency, conciseness, and aesthetics (Xuan et al., 9 May 2026).

The headline results are as follows.

Setting	Metric	Result
HPR-Benchtag after SFT	Avg@10	0.7325
Full UserGPT on HPR-Benchtag	Avg@10	0.7306
Full UserGPT on HPR-Benchsum	AccEx	0.7528
Full UserGPT on HPR-Benchsum	COVEx	0.9747
Full UserGPT on HPR-Benchsum	BLEU-4	0.0691
Full UserGPT on HPR-Benchsum	Scoresim	0.8215

These figures are reported in the technical report (Xuan et al., 9 May 2026). The paper further states that UserGPT is nearly identical to Qwen3.6-Plus (0.7329) on HPR-Benchtag despite using only 8B parameters, and that on HPR-Benchsum it exceeds Qwen3-235B-A22B-Thinking on AccEx and COVEx, specifically 0.7528 vs 0.7014 and 0.9747 vs 0.9638 respectively (Xuan et al., 9 May 2026).

The ablations are central to the article’s argument. For controversial-data training, HPR-Benchtag reaches Pass@1 = 0.7316, Pass@5 = 0.8922, and Pass@10 = 0.9259, which the report interprets as evidence that difficult samples teach reasoning rather than memorization (Xuan et al., 9 May 2026). In the curriculum ablation, the full curriculum yields AccEx = 0.6899 and COVEx = 0.9587, outperforming partial-stage variants (Xuan et al., 9 May 2026). In the RL ablation, the sequence is:

• SFT: AccEx 0.6899, COVEx 0.9587
• GRPO: AccEx 0.7046
• GRPO + group-level filtering: AccEx 0.7323, COVEx 0.9681
• DF-GRPO: AccEx 0.7528, COVEx 0.9747

This progression is used to support the claim that both filters are important and that the dual-filter version is best overall (Xuan et al., 9 May 2026).

The report also states that specialization does not catastrophically damage general ability: on C-Eval, CMNLI, BUSTM, C3, and ChineseEcomQA, the model suffers only a 1.32-point drop relative to the backbone’s thinking variant while outperforming the non-thinking backbone by 5.95 points (Xuan et al., 9 May 2026).

7. Interpretation, limitations, and relation to adjacent work

The report presents UserGPT’s strengths through qualitative examples in which the model infers a family-oriented female user from purchases such as diapers, children’s clothing, school supplies, and pet-related items, then produces both a structured JSON profile and a coherent narrative summary (Xuan et al., 9 May 2026). The strengths emphasized are consistent inference from multiple weak signals, integration of cross-year evidence, recognition of life-stage transitions, and concise persona summaries.

The limitations are stated explicitly. The report notes continued difficulty with deeper implicit and temporal reasoning, hallucination suppression, effective use of longer contexts, and multimodal user profiling with image/video signals (Xuan et al., 9 May 2026). It also lists recurring failure modes, even for frontier LLMs: temporal confusion, over-interpreting one-off behaviors, mixing personas in shared-account settings, and failure to distinguish personal preference from external or accidental actions (Xuan et al., 9 May 2026).

A notable future direction is Incremental Profiling, formulated as

$\hat{y}^s$1

which would update persona memory recursively rather than recomputing from scratch (Xuan et al., 9 May 2026). This suggests a persistent-memory architecture for personalized agents.

In the broader literature, BehaveGPT offers a related but distinct paradigm: a transformer-based model for large-scale user behavior prediction using a DRO-based pretraining paradigm, evaluated on next behavior prediction, new behavior prediction, long-term generation, and cross-domain adaptation (Gong et al., 23 May 2025). BehaveGPT reports more than 600 million behavior logs, a scaling-law fit

$\hat{y}^s$2

with $\hat{y}^s$3 and $\hat{y}^s$4, and more than a 10% improvement in macro and weighted recall in its domain (Gong et al., 23 May 2025). This adjacent line of work supports the general proposition that user behavior supports foundation-model-style pretraining, while UserGPT extends that direction toward coherent persona summarization rather than only predictive sequence modeling.

Taken together, these results position UserGPT as a generative framework for converting fragmented digital traces into a compressed and evolving persona memory (Xuan et al., 9 May 2026). A plausible implication is that its main novelty lies not in any single component, but in the way simulation, semantization, curriculum design, and RL alignment are combined to make long-horizon persona reasoning tractable for an 8B model.