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

AI summary

Overview

Research area: Recommender systems evaluation, LLM-based user simulation, and psychological user modeling (submitted under Natural Language Processing).

Technical level: Advanced. The paper assumes familiarity with LLM agents, latent-state transition models, optimal transport (Gromov–Wasserstein), and standard recommender benchmarks.

Scope: The paper proposes a psychology-inspired simulated user (PsyBer-Agent) with an explicit three-dimensional latent-state engine, aligns it to real interaction logs via Unbalanced Gromov–Wasserstein transport, and introduces the WebSim platform and PsyBer Benchmark for closed-loop recommender evaluation.

What This Paper Is About

Testing a recommender system before launch normally requires real users, which is slow, costly, and risky, so researchers build LLM-based "simulated users" instead. The paper argues that existing simulators are surface imitators: they produce plausible short-term clicks but have no explicit mechanism for how users get tired, satiated, emotionally satisfied, or lose self-congruence over long sessions, so errors accumulate into behavioral drift. The goal is a simulator whose internal state changes in a controllable, inspectable way and whose behavior can be calibrated against real click logs without fine-tuning a large backbone model.

Key Contributions

1. PsyBer-Agent, a psychology-inspired user simulator built on an explicit, model-defined latent-state engine (Psy-Engine) that tracks three evolving coordinates in [0,1]: cognitive reserve (C), emotional satisfaction (E), and self-congruence (S), updated by a bounded relaxation rule with rate α = 0.35.
2. A psycho-behavioral alignment framework that replays logged clicks through Psy-Engine to build reference latent trajectories, matches them to simulator trajectories with Unbalanced Gromov–Wasserstein (UGW) optimal transport, and distills the resulting transport plan into soft action targets for a lightweight "fast policy" — no backbone LLM fine-tuning required.
3. WebSim, an end-to-end Flask-based platform for multimodal asset preparation, offline alignment/policy training, and browser-level interactive simulation, with decoupled user interaction and recommender inference.
4. The PsyBer Benchmark, spanning multimodal scenarios (MM-ML-1M for movies, Amazon Review for shopping, MicroLens and NineRec for image-text feeds and short video) and seven recommender backbones (PopRec, BPR-MF, LightGCN, Multi-VAE, GRU4Rec, SASRec, BERT4Rec), with evaluation of behavioral fidelity, latent-state consistency, robustness, and social-cue perturbations.

Main Findings

• Close agreement with human users under a fixed backend: comparing 500 human users against 500 simulated users on the same interface, the full PsyBer-Agent with Qwen3-VL-235B differed from humans by 0.0030 HR@10 (0.2350 vs 0.2380), 0.0016 NDCG@10 (0.1150 vs 0.1166), and 0.0040 CTR (0.2960 vs 0.3000). Removing memory widened the gaps to 0.0160, 0.0106, and 0.0220; also removing the user profile widened them to 0.0400, 0.0246, and 0.0500.
• Fidelity varies by scenario more than by persona: Amazon Magazine Subscriptions produced compact, favorable clusters (JSD 0.2085–0.2173, Pearson R 0.5164–0.5500, EMD 0.2245–0.2491 across backbones), while MM-ML-1M was more dispersed (JSD 0.3249–0.6515, R 0.1217–0.1867, EMD 0.1649–0.4602). Identity-gap and value-oriented profile groups achieved the highest composite fidelity.
• Best overall against simulator baselines: in a fixed-candidate replay on MM-ML-1M (500 users, 10 decision steps, five seeds), PsyBer-Agent reached 0.275 accuracy, 0.406 CTRGap, 0.286 JSD, 0.305 Pearson R, and 2.53s latency, versus DyTA4Rec (0.254 / 0.455 / 0.335 / 0.281 / 5.91s), RecAgent (0.212 / 0.482 / 0.420 / 0.264 / 5.46s), Agent4Rec (0.193 / 0.489 / 0.447 / 0.264 / 5.74s), and UserSim (0.128 / 0.500 / 0.605 / 0.142 / 3.31s).
• Scales to 5M queued tasks: across N ∈ {500, 1K, 100K, 1M, 5M}, p95 latency per decision grew only from 2.52s to 3.80s, throughput stabilized near 4.1 tasks/s for N ≥ 100K (4.18, 4.12, 4.09), success rates were 100% at 500 and 1K, 92.75% at 100K, 91.34% at 1M, and 95.18% at 5M, with wall-clock rising from 8.3 minutes to 20,358.9 minutes and token cost from 4,262.7K to 52,262.77M.
• Every component contributes: the full model reached 0.2608 CTRGap, 0.4297 Exp-JS, and 0.0636 CES-MSE, versus 0.8926 / 0.7852 / 0.1654 for a direct prompt-only agent; removing Psy-Engine (0.7469 / 0.7382 / 0.1154), optimal transport (0.6906 / 0.6817 / 0.1053), preprocessing (0.8652 / 0.7613 / 0.1290), or fusion memory (0.2985 / 0.6764 / 0.1254) degraded at least one metric.
• Information channels are complementary: dropping scene information (0.5243 CTRGap), user identity profile (0.6738), item information (0.7824), or psychological-change signals (0.8816) each hurt relative to the full 0.2608.
• Deployment-scale run statistics: across 200 runs and 3,981 decisions the system showed 3.11s average latency, full fast-hint coverage, and a 99.9% batch success rate.
• Stated limits: CES-MSE measures consistency inside the authors' own model-defined state space, not validated correspondence to measured human psychology; the human–simulator study does not establish satisfaction prediction or live A/B ranking of competing recommenders.

Methodology in Plain English

The researchers give each simulated user three numbers between 0 and 1: how much cognitive "room" they have left, how emotionally satisfied they are, and how much an item fits their self-image. Item features (semantic embedding, identity gap, cognitive load, social proof) feed fixed, bounded rules that compute an attractiveness signal, a fatigue signal, and target values for satisfaction and self-congruence. Each step, the three numbers move partway toward those signals — a simple, bounded update rather than free-form LLM role-play.

Because real logs contain clicks but never these internal numbers, the team replays each user's logged click sequence through the same engine, treating logged items as forced inputs. This creates a pseudo-expert trajectory of latent states. They then match this reference trajectory to the states produced by a live simulator using Unbalanced Gromov–Wasserstein transport, which tolerates trajectories of different lengths and rhythms. The resulting matching weights are converted into soft action labels over each candidate slate (with leftover mass assigned to "skip"), and a small fast policy is trained on those labels. At inference the fast policy proposes the action, and the backbone LLM or vision-language model is called only when the policy is unsure or the action space is unsupported; random selection is the last resort. Evaluation uses WebSim, a Flask-based loop with seven recommender backbones, chronological 70/10/20 splits, and metrics for distribution mismatch (JSD), item-level agreement (Pearson R), post-prefix behavioral discrepancy (EMD), CTR calibration (CTRGap), and latent-state consistency (CES-MSE).

Why This Matters

Impact on research: The paper shifts user simulation from prompt engineering toward explicit state-transition models that can be inspected, ablated, and scaled, and it offers a reproducible testbed (WebSim plus the PsyBer Benchmark) with seven backbones and multiple domains. Its negative framing — that latent states are computational constructs, not validated psychological measurements — gives other researchers a clear boundary to test against.

Real-world applications:

• Pre-deployment sanity checks of a new recommender before running costly online A/B tests.
• Robustness and stress testing against social-cue perturbations and manipulative recommendation strategies.
• Multi-domain prototyping where real traffic is scarce, such as short-video feeds, image-text information flows, and niche e-commerce catalogs.
• Reproducible, privacy-preserving behavioral studies using simulated cohorts instead of live users.

Industry relevance: The scalability results — up to 5M queued simulation tasks under bounded concurrency, p95 latency under 3.80s per decision, and explicit token-cost accounting up to 52,262.77M tokens — directly address whether simulated users are affordable at industrial scale, and the fast-policy design keeps the expensive backbone calls off the critical path.

Future Directions

• Adding independently validated supervision, so latent states can be checked against measured psychological or survey data rather than only internal consistency.
• Extending evaluation across more domains, longer horizons, and additional robustness settings beyond the tested fixed-replay protocol.
• Determining whether simulator agreement under a fixed backend translates into correct ranking of competing recommenders in a live A/B test, which the authors explicitly do not claim.
• Improving the scenarios where fidelity was weakest — notably the semantically diverse movie-consumption setting (MM-ML-1M) relative to compact e-commerce catalogs — and tightening success rates at the 100K and 1M scales, where they were 92.75% and 91.34%.

Target Audience

Recommender-system researchers and engineers who need credible offline evaluation; practitioners building LLM agent simulators or synthetic user populations; and NLP/agent researchers interested in aligning generative agents to logged behavior with optimal-transport objectives. Readers without background in optimal transport or latent-variable modeling will find the framework sections demanding, though the experimental results are readable on their own.

Authors’ abstract

Evaluating recommender systems before deployment requires simulators that remain realistic over long, closed-loop interaction cycles. Current LLM-based agents generally function as surface imitators that generate realistic short-term responses but fail to capture the underlying user dynamics such as fatigue, affect, and shifting preferences. We propose PsyBer-Agent, a psychology-inspired simulator that models these dynamics through its Psy-Engine. This engine represents each user with three evolving, model-defined latent states. To calibrate simulated behavior against real interaction logs, we align log-conditioned reference trajectories with simulator-generated latent-state trajectories via Unbalanced Gromov--Wasserstein optimal transport. This approach provides structured supervision for a lightweight policy without the computational burden of fine-tuning the backbone LLM. We also introduce WebSim and the PsyBer Benchmark, which encompass multimodal scenarios and seven recommender backbones. Our framework evaluates behavioral fidelity, latent-state consistency, robustness, and responses to social-cue perturbations. On the tested horizons, PsyBer-Agent shows lower behavioral discrepancy than the prompt-only simulators while maintaining stable performance under large-scale concurrent simulation for up to 5M queued simulation tasks under bounded concurrency. Code is available in the supplementary material and at: PsyBer-Agent: https://anonymous.4open.science/r/CC66F/, WebSim Platform: https://anonymous.4open.science/r/D8EAX/.