PsyBer-Agent: A Psychology-Driven User Agent for Recommender Systems

1 Introduction

• Surveys, user studies, and online A/B tests are costly, slow, and risky, while offline log replay misses postdeployment counterfactual interaction dynamics.LLM-based agents can generate click, skip, and continued-browsing trajectories as a richer testbed for recommender evaluation, A/B testing, and robustness analysis.Source: §1 · p. 2
• PsyBer-Agent models cognition, emotional/social satisfaction, and self-congruence as evolving states, then uses UGW alignment to distill offline expert trajectories into a lightweight policy without backbone fine-tuning.Its states interact with personas, memory, and multimodal item representations to generate behavior through explicit transitions rather than prompt-only role-playing.Source: §1 · p. 2
• Across WebSim evaluations, PsyBer-Agent more faithfully matches real interaction logs, better matches model-defined latent-state trajectories, and remains stable under perturbations and stress.The platform supports up to 5M queued simulation tasks under bounded concurrency, with bounded latency growth, high success rates, and explicit cost accounting.Source: §1 · p. 2
• The introduction frames the contribution as a simulator, alignment framework, and benchmark/platform for closed-loop testing, behavior fidelity, robustness, and million-scale simulation.Table 1 compares agent-based simulators on dynamic state modeling, psychological modeling, and real-log alignment.Source: Table 1 · p. 2, §1 · p. 3

2 Problem Formulation and Notation

• Each implicit log records a user, an interacted item, and its interaction timestamp, and the user’s expert sequence is reconstructed from those records.The log is written as D = {(u, i, t)}.Source: §2 · p. 3
• At each step, the simulator observes a top-K slate and selects one of its items or skip.The available action set is {1, …, K, skip}.Source: §2 · p. 3
• Items carry semantic, identity-gap, cognitive-load, and social-proof features, while each user has cognition, emotional/social satisfaction, and self-congruence states plus persona and memory.The psychological state is represented as a three-coordinate vector in [0, 1].Source: §2 · p. 3
• The formulation connects the simulator’s state and action setup to the goal of aligning logged behavior with model-defined state transitions.The supplied notation also indicates a fast policy trained through UGW alignment to obtain soft action targets.Source: §2 · p. 3
• The central question is whether aligning observed user behaviors with psychology-inspired, model-defined state transitions can yield a reliable behavioral simulator for recommendation systems.The formulation starts from implicit interaction logs.Source: §2 · p. 3

3 PsyBer-Agent Framework

• The framework couples Psy-Engine state definitions, transitions, and action evaluation with multimodal item features, reconstructed log-conditioned reference trajectories, and UGW-based fast-policy learning.WebSim executes browser interactions and records traces, states, and run summaries for reproducible evaluation.Source: §3 · p. 3, §3.1 · p. 3
• The C/E/S states are computational constructs motivated by psychological principles, so CES-MSE measures internal consistency rather than validated correspondence to independently measured human states.The anchors condition the simulator but are not interpreted as direct measurements of psychological traits.Source: §3 · p. 3, §3.1 · p. 4

3.1 Simulation Experiment Platform Construction

• WebSim prepares multimodal assets and runs browser-level simulations using a trained fast policy when available or a constrained vision-language fallback otherwise.Its browser controller executes click and pagination actions and logs traces, psychological states, and run summaries.Source: §3.1 · p. 3

3.2 Psy-Engine Modeling

• Psy-Engine maintains bounded cognition, emotional-satisfaction, and self-congruence states as model-defined variables motivated by psychological concepts.Item signals are organized into attractiveness and fatigue proxies and interaction-dependent targets; these are fixed, bounded computational rules rather than uniquely determined psychological laws.Source: §3.2 · p. 4
• The bounded update moves cognition according to net attractiveness minus fatigue, while emotional satisfaction and self-congruence move toward interaction-dependent targets.A shared relaxation rate keeps the parameterization compact rather than assuming identical psychological time constants; α is set to 0.35.Source: §3.2 · p. 5

3.3 Psycho-behavioral Alignment

• Logged clicks are replayed as forced inputs to Psy-Engine, producing model-defined reference states for alignment rather than observed psychological annotations.The reconstructed reference trajectory and simulator trajectory can differ in length and browsing rhythm, so alignment begins with a soft state prior derived from expert prefixes.Source: §3.3 · p. 5
• UGW aligns simulator and expert trajectories, and the resulting transport plan supplies soft action targets for a lightweight fast policy.Matched expert items map to corresponding slate actions, while residual unmatched transport mass is assigned to skip; the backbone LLM is invoked only when needed or when policy confidence is insufficient.Source: §3.3 · p. 6

4.1 Benchmark Coverage

• The PsyBer Benchmark defines two core tasks: multi-scenario online user simulation and recommendation-model performance evaluation.Figure 2 summarizes dataset footprint, scale and density, and scenario and backbone coverage.Source: §4.1 · p. 6, Figure 2 · p. 7

4.2 Task Evaluation

• WebSim decouples user interaction from recommender inference and uses random fallback for unavailable or empty model outputs to support consistent cross-model and cross-dataset evaluation.The platform is Flask-based and operates as a closed-loop simulator.Source: §4.2 · p. 6
• The benchmark includes representative popularity, matrix-factorization, graph, variational, sequential, and transformer recommenders, alongside movie, shopping, image-text, and short-video scenarios.Named backbones include PopRec, BPR-MF, LightGCN, Multi-VAE, GRU4Rec, SASRec, and BERT4Rec; datasets include MM-ML-1M, Amazon Review, MicroLens, and NineRec.Source: §4.2 · pp. 6–7
• Evaluation covers online A/B tests, behavior fidelity, baseline simulator comparison, and scalability and cost.The benchmark’s stated scope includes multiple recommendation backbones and application scenarios.Source: §4.2 · p. 7

5.1 Experimental Settings

• The evaluation uses representative recommender backbones and agent-based simulator baselines, with chronological 70/10/20 train/validation/test splits and held-out users.The experiment plan includes population- and individual-level fidelity, baseline comparison, scalability and token cost, component and information ablations, and an offline backbone sanity check.Source: §5.1 · p. 7, §5.2 · p. 7

5.2 Evaluation and Analysis

• With Qwen3-VL-235B, full-model differences from human users are 0.0030 for HR10, 0.0016 for NDCG10, and 0.0040 for CTR.Removing memory increases these differences to 0.0160, 0.0106, and 0.0220; removing memory and profile increases them to 0.0400, 0.0246, and 0.0500.Source: §5.2 · p. 8, Table 2 · p. 8
• Across scenarios and backbones, Amazon Magazine Subscriptions has more compact low-JSD/high-R/low-EMD clusters than the more dispersed MM-ML-1M results.The metrics capture population-distribution mismatch, item-level agreement, and downstream discrepancy after matched prefixes, respectively.Source: Table 3 · p. 9, §5.2 · p. 9
• Identity-gap and value-oriented groups achieve the highest composite fidelity scores, while remaining mismatch appears more related to scenario and recommender effects than unstable persona conditioning.The Figure 3 score combines normalized JSD, Pearson R, and EMD.Source: Figure 3 · p. 8, §5.2 · p. 9
• PsyBer-Agent achieves the best overall performance in controlled replay, with highest accuracy and correlation and lowest CTRGap, JSD, and inference latency among methods.The comparison used MM-ML-1M, 500 users, 10 decision steps per user, and five random seeds.Source: Table 4 · p. 9, §5.2 · p. 9
• At 5M tasks, p95 latency is 3.80s and success rates remain above 91%; workloads of at least 100K reach approximately 4.1 tasks/s after queue warm-up.Across the tested workloads, p95 latency grows from 2.52s to 3.80s, with wall-clock time and token cost also reported.Source: §5.2 · pp. 9–10, Table 5 · p. 10

5.3.1 Component Ablation

• The full component configuration has the lowest CTRGap, Exp-JS, and CES-MSE; removing preprocessing, Psy-Engine, or transport alignment increases one or more discrepancies.Removing fusion memory has a smaller CTRGap effect than several other removals but increases CES-MSE; the results do not establish psychological validity of the latent states.Source: Table 6 · p. 10, §5.3.1 · p. 10

5.3.2 Information Ablation

• The full information-channel configuration performs best overall, and removing any individual channel increases at least one reported discrepancy.The tested channels include item text and visuals, psychological-change signals, user identity profiles, and scene information; the ablations measure input sensitivity rather than validate latent traits.Source: Table 7 · p. 10, §5.3.2 · p. 10

User Simulation for Recommender Evaluation

• Recent recommender simulators use persona and working-memory architectures, feedback distillation, latent reasoning signals, or human-camouflaged adversarial users.Long-horizon agent benchmarks also examine memory across tool-use, research, coding, and web tasks, while MMRC reports memory degradation and error accumulation in sustained multimodal conversations.Source: §6 · p. 11

Psychological User Modeling and Behavioral Calibration

• Psychological and calibration-oriented modeling spans Big Five profiles, cognitive alignment, values and motivations, co-evolving feedback, and behavior-level calibration across simulation domains.Examples include mobility, economic-expectation, public-health, investor, and online-discourse simulations.Source: §6 · p. 11

Robustness, Safety, and Large-scale Simulation

• Reported risks in LLM-centered recommendation and agent simulation include poisoning, backdoors, privacy leakage, prompt injection, manipulation, shilling, and preference-oriented exploits.These risks motivate evaluation that includes fidelity, scalability, robustness, privacy-aware analysis, and controlled pre-deployment diagnosis.Source: §6 · p. 11

• PsyBer-Agent combines model-defined latent-state dynamics, offline UGW alignment, and a fast policy with LLM/VLM fallback for closed-loop recommender evaluation.Source: §7 · p. 11
• Across 200 runs and 3,981 decisions, PsyBer-Agent achieves 3.11 s average latency, full fast-hint coverage, and 99.9% batch success.On the PsyBer Benchmark, it reduces behavioral discrepancy and improves model-defined state consistency over prompt-based baselines under fixed replay.Source: §7 · p. 11
• The results do not establish psychological validity or predict live A/B rankings; future work will broaden evaluation across domains, horizons, and robustness settings.Source: §7 · p. 11

AI use statement

• Generative AI assisted with language polishing, editorial work, and WebSim and experiment-script coding; authors reviewed and tested its outputs and take responsibility for the work.The research ideas, methodology, experimental design, and reported results were produced and validated by the authors.Source: AI use statement · p. 12

Ethics Statement; Reproducibility statement

• The human–simulator agreement study involved 500 consenting participants, received institutional ethics approval, and stored interaction logs anonymously; PsyBer-Agent is intended as a diagnostic proxy.The paper provides implementation, metric, hardware, and protocol details, and links to source code for PsyBer-Agent and WebSim.Source: Ethics Statement · p. 12, Reproducibility statement · p. 12

A.4 Intermediate Drives; A.5 State Transition Equations

• Psy-Engine represents cognition, emotional satisfaction, and self-congruence, using attention and fatigue drives plus interaction-dependent targets to update its model-defined states.The state-transition design is a computational hypothesis, not a claim of independently established psychological validity.Source: §A.1 · p. 21, §A.4 · p. 23, §A.5 · p. 23

A.2 Cross-domain feature construction

• Cross-domain item features combine text and visual embeddings with preference incongruence, content complexity, and social-proof signals, normalized using pre-window information.The three persona anchors represent static profile, long-term preference, and recent preference; item-level signals are kept stable across interaction steps.Source: §A.2 · pp. 21–22

B.1 Model-Defined State Sequence Construction

• Because clicks do not identify latent psychology, logged expert actions are replayed through the same persona-conditioned Psy-Engine to construct model-defined reference states, not observed psychological labels.Replay uses teacher forcing and bounded memory to align expert-side and simulator-side transition conditions.Source: §B.1 · p. 24

B.2 Transport Initialization and UGW Alignment

• An offline regression prior initializes Unbalanced Gromov-Wasserstein alignment, whose relaxed transport plan accommodates differing sequence lengths, browsing speeds, and unmatched segments.The alignment compares expert-item semantic structure with simulator-state structure and penalizes structural discrepancies.Source: §B.2 · pp. 25–26

B.3 Alignment-induced Soft Supervision for Policy Learning

• The optimized alignment yields state-conditioned soft action targets projected onto the current slate, assigning residual mass to skip for unavailable or unmatched expert items.These dense targets train the lightweight online policy with cross-entropy rather than relying only on sparse click replay.Source: §B.3 · pp. 26–27

C.1 Benchmark Overview

• The PsyBer Benchmark standardizes datasets, recommendation scenarios, recommender backbones, and evaluation protocols for closed-loop user-simulator evaluation on WebSim.WebSim decouples user-interaction serving from recommender inference to support consistent comparisons across models and datasets.Source: §C.1 · p. 27

C.2 Data Construction

• Seven public datasets span movie, e-commerce, short-video, and cross-domain recommendation, including MM-ML-1M Zhang et al. (2026d), Amazon subsets Ni et al. (2019), MicroLens-100K Ni et al. (2025), and NineRec-Bili500K Zhang et al. (2025a).All datasets use standardized interaction formatting, scenario-specific item metadata, and chronological training and validation/test splits.Source: §C.2 · p. 27

C.3 Task Definition

• The benchmark defines multi-scenario closed-loop user simulation and recommendation-model evaluation as its two core tasks.Simulation examines feedback trajectories such as clicks, skips, and continued browsing; model evaluation uses simulated-user trajectories to measure recommendation quality and system behavior, including CTR, NDCG@K, CTRGap, and latency per decision.Source: §C.3 · pp. 27–28

C.4 Benchmark Coverage and Evaluation Protocol

• Coverage spans seven recommender families and movie, shopping, image-text information-flow, and short-video scenarios, evaluated through a unified closed-loop protocol.WebSim serves recommendation slates, collects simulator feedback, and updates interaction history stepwise; evaluation considers behavior fidelity, baseline comparison, scalability and cost, and group behavior.Source: §C.4 · p. 28

C.5 WebSim Experiment Platform

• WebSim separates interaction serving, model-agnostic recommendation inference, and dataset-aware metadata and poster delivery, with a wrapper dispatching to available recommender backbones.The wrapper normalizes item IDs, filters unavailable or duplicate candidates, and falls back to a default model or random sampling when inference is unavailable or yields no valid items.Source: §C.5 · pp. 28–29
• WebSim manages per-session history, recommendation caches, and interaction statistics, while dataset-aware asset handling and configurable launch paths support serving and reproducible evaluation.Sessions reset with four seed-controlled items; history is capped at 20, page size is 4, and the recommendation cache uses top-K = 200. Missing posters receive generated SVG placeholders.Source: §C.5 · pp. 28–29

System architecture

• WebSim uses a browser-based grid interface for click and pagination feedback and a Flask backend for session, recommendation, poster, cleanup, and health services.Figure 4 presents the platform’s UML class diagram.Source: System architecture · p. 29, Figure 4 · p. 29

C.6 Metrics and Evaluation

• Behavior fidelity is measured at population and trajectory levels using Interaction JSD (lower means closer population interaction patterns), EMD (lower means closer future trajectories after matched prefixes), and step-wise Decision Accuracy.Source: §C.6 · p. 30
• CES-MSE compares simulator-generated latent states with reference states reconstructed from observed interactions, measuring internal consistency rather than validated human psychological states.A lower value does not establish correspondence to independently measured psychological states.Source: §C.6 · pp. 31–32
• Robustness measures compare perturbed with clean behavior and decisions, while social-cue diagnostics track sentiment conformity, herding, polarization, and visual-salience associations.Comment conformity is a manipulation check, not standalone evidence of action-level herding.Source: §C.6 · pp. 31–32

D.1 Details of the Evaluated Datasets

• The PsyBer Benchmark covers seven multimodal datasets across movie, e-commerce, short-video, and cross-domain settings, including MicroLens-100K Ni et al. (2025) and NineRec-Bili500K Zhang et al. (2025a).NineRec-Bili500K is used for transferable and cross-domain recommendation robustness evaluation.Source: §D.1 · pp. 32–33

D.2 Details of the Evaluated Recommender Systems

• Evaluation spans seven recommender backbones across five paradigms: heuristic, collaborative filtering, graph neural network, deep generative, and sequential recommendation.The model list is summarized in Table 9.Source: §D.2 · p. 33, Table 9 · p. 33

D.3 Details of the User Simulator Baselines

• Baseline simulators enter quantitative comparisons only when executable artifacts, required checkpoints, and a protocol-compatible action interface can be verified.The excluded methods are discussed as related work, not claimed to be unavailable or inferior.Source: §D.3 · p. 34

D.4 Details of the Prompt Design

• Final action selection combines a live WebSim screenshot with psychological state, fast-policy suggestions, recent clicks, and available page items; the prompt constrains outputs to valid candidate actions.State and policy suggestions serve as reference signals rather than hard constraints.Source: §D.4 · pp. 34–35, Figure 5 · p. 35

D.5 Details of Experimental Hardware Environment and Parallel Simulation Settings

• A queue-based launcher assigns independent user processes to bounded asynchronous workers and uses API-pool slots for LLM/VLM decisions.The experiments use a dedicated server with 80 CPU cores, 251 GB memory, and six RTX 3090 GPUs.Source: §D.5 · p. 36
• Under a fixed configuration, increasing workload from 100K to 1M and 1M to 5M increases token consumption by 10.45× and 5.09×, and execution time by 10.19× and 5.02×, indicating approximately linear cost growth.Source: §D.5 · p. 36
• Simulation cost grows approximately with tasks multiplied by interaction rounds; the reported unit cost is about $4.45 per 1K tasks, so 5M-task runs are framed as stress tests rather than routine evaluation.The reported token total includes offline trajectory construction and training, environment calls, and online LLM/VLM fallbacks.Source: §D.5 · p. 36
• LLM/VLM fallback handles approximately 0.6% of decisions and random fallback approximately 0.04%, leaving the fast policy to handle the overwhelming majority.The authors state that fidelity is not dominated by fallback behavior.Source: §D.5 · p. 36

D.6 Details of Fixed-Backend Human–Simulator Agreement Protocol

• The human–simulator study compares 500 real users with 500 simulated users over 10 browsing rounds, holding the interface and backend fixed and treating repeated decisions as clustered within user.It uses a prespecified user-level paired bootstrap for 95% intervals when user-level outcomes are available.Source: §D.6 · pp. 36–37

E.1 PsyBer-Agent Group Behavior Exploration.

• Under social-cue perturbations, PsyBer-Agent reflects comment sentiment, but this uptake does not translate into strong action-level herding, polarization, or visual-salience-driven behavior.The experiment fixes the candidate pool and base ranking while perturbing rating priors, hotness badges, and comment sentiment.Source: Table 11 · p. 37, §E.1 · p. 37

Update and Training Hyperparameters

• The hyperparameter analysis sweeps four to five values per parameter around a fixed configuration, with five random-seed repetitions per setting.Table 12 reports mean CES-MSE and latency per decision.Source: §E.2 · p. 37, Table 12 · p. 38
• Within the tested sweeps, α = 0.35 achieves the lowest mean CES-MSE but higher latency, while lr = 10^-2 has slightly lower CES-MSE and substantially higher latency than 3 × 10^-3.The chosen defaults are α = 0.35 and lr = 3 × 10^-3.Source: §E.2 · p. 37, Table 12 · p. 38

Aggregation-Weight Sensitivity

• The aggregation analysis prespecifies eight configurations, varies one aggregation function at a time, and holds the state-update equations and remaining hyperparameters fixed.Each variant regenerates reference trajectories and retrains its regression prior and fast policy using identical splits, budgets, and seed sets.Source: Table 13 · p. 38, §E.2 · p. 38
• Removing Psy-Engine changes reported mean metrics more than tested weight adjustments; the default reduces CTRGap and JSD by 0.050 each and improves accuracy by 0.035.No configuration is best on every metric, and the results support equal weighting as a default rather than establish its optimality.Source: §E.2 · pp. 38–39, Table 14 · p. 39
• Across tested aggregation weights, maximum absolute changes from default are 0.010 in CTRGap, 0.015 in Interaction JSD, and 0.007 in Decision Accuracy.The evaluation uses MM-ML-1M, a fixed recommender backbone within each scenario, and five seeds per configuration.Source: §E.2 · p. 38, Table 14 · p. 39

E.3 PsyBer-Agent Robustness Stress Analysis

• Under tested stresses, local perturbations and memory noise were tolerated, but longer horizons, domain shift, and conflicting signals degraded behavior or decisions.Relative Drop rose to 55.18% at track=100; memory perturbations retained a 1.00 success rate with latency of 4.35s–4.51s versus 4.19s clean.Source: §E.3 · pp. 40–41

E.4 Representative Benchmark Performance of Recommendation Models on the Test Platform

• Near-ceiling results on some sparse shopping subsets indicate these settings are comparatively easy for strong sequential recommenders and primarily check WebSim’s backend layer.The representative model table is described as a backend sanity check, not the primary benchmark result.Source: §E.4 · p. 41

E.5 Profile-Conditioned Group Fidelity Diagnostics

• Across profile groups, Amazon Magazine Subscriptions had lower JSD and EMD than MM-ML-1M, while movie profiles varied more and novelty-seeking users showed higher JSD.The diagnostics indicate simulator-side profile differences anchored to aggregate interaction logs, not direct proof of real-world persona-level fidelity.Source: §E.5 · pp. 41–42

E.6 Offline Backbone Sanity Check

• On the offline split, BPR-MF had the highest point estimates, followed by LightGCN and SASRec; this ordering checks the backend, not online recommender ranking validity.The check used MovieLens-1M logs, 500 users, and held-out last interactions; it provides neither human outcomes nor human–simulator ranking validation.Source: §E.6 · p. 43

E.7 Aggregate Baseline Comparison

• The main-text Table 4 baseline comparison is the primary quantitative result, based on 500 users over 10 steps per user and five random seeds.The appendix provides no separate estimates for individual backbones.Source: §E.7 · pp. 43–44

E.8 Fallback Mechanisms and Usage Statistics

• Most observed decisions were handled by the fast policy, while random fallback contributed only a small fraction of actions in the reported workload.Fallback activation rates are reported over all decision steps.Source: §E.8 · p. 44

E.9 Large-Scale Cost and Execution Analysis

• Scaling from 100K to 1M tasks raised token use and execution time about 10.45× and 10.19×; from 1M to 5M, increases were about 5.09× and 5.02×.The estimated API cost was approximately $4.45 per 1K tasks; 5M denotes queued tasks under the reported configuration, not simultaneous active tasks, and is a stress test rather than a universal cost-effectiveness claim.Source: §E.9 · p. 44
• Prior-guided UGW achieved accuracy 0.278 and CTRGap 0.406, with similar latency across alignment variants (2.48–2.53s/decision).Relative-position matching to uniform-initialized UGW improved accuracy by 1.4pp and reduced CTRGap by 0.023; prior guidance added 1.1pp and reduced CTRGap by 0.012.Source: §E.9 · p. 45

F.1 MovieLens-based Web Browsing Scenario

• In MovieLens, C, E, and S generally declined with browsing, with self-congruence falling fastest; selected Top-30 popularity broadly followed offline data except for higher Toy Story selection.The scenario uses a MovieLens browsing interface with LightGCN; the offline regression prior initializes psycho-behavioral alignment, then UGW alignment optimizes the fast policy.Source: §F.1 · p. 46, Figure 9 · p. 47, Figure 10 · p. 48

F.2 Amazon Magazine Subscriptions Shopping Scenario

• In Amazon Magazine Subscriptions, psychological factors attenuated over browsing and responded to lower relevance, while Top-30 selections broadly matched offline head-item popularity with selective deviations.The scenario uses GRU4Rec; an offline regression prior initializes alignment before UGW optimizes the fast policy for clicks, skips, and pagination.Source: §F.2 · p. 47

G.1 Limitations

• The three-dimensional cognition, emotion, and self-congruence state model may omit other drivers, while pseudo-expert trajectories reuse Psy-Engine rather than independently validated human psychological measurements.Fidelity may also degrade at substantially longer horizons, in unseen domains, or in interfaces with interaction patterns unlike WebSim.Source: §G.1 · p. 48
• Simulated users cannot fully replace real-user studies or online A/B tests, and large-scale simulation partly depends on external LLM APIs and public interaction logs.API-version reproducibility, privacy protection, and fairness across user groups remain practical considerations.Source: §G.1 · pp. 48–49

G.2 Broader impacts

• The platform may support pre-deployment diagnosis, but psychology-driven simulation could be misused for manipulative recommendations or raise privacy concerns; the authors advise human oversight and staged validation.The paper also recommends dataset provenance documentation and checks for distributional bias, herding, and fairness across user groups.Source: §G.2 · pp. 49–50