acceptodds
Under review as a conference paper at ICLR 2027

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

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 , a psychology-inspired simulator that models these dynamics through its . 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: https://anonymous.4open.science/r/CC66F/ , https://anonymous.4open.science/r/D8EAX/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Recommender Systems, User Simulation, LLM-based Agents, Long-term User Behavior

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kevvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PsyBer-Agent, a user simulator for pre-deployment recommender evaluation. Its main components are: 1) Psy-Engine, which maintains three evolving, model-defined latent states: cognitive reserve, emotional satisfaction, and self-congruence. State transitions incorporate user personas, memory, and item features. 2) A psycho-behavioral alignment framework that replays logged interactions through Psy-Engine to construct model-defined reference trajectories and uses Unbalanced Gromov–Wasserstein (UGW) alignment to derive soft action supervision. 3) A lightweight fast policy that handles most decisions, with LLM/VLM fallback when needed.

WebSim and the PsyBer Benchmark, are proposed to evaluate behavioral fidelity, model-defined latent-state consistency, robustness, and scalability, by comparing the simulated user behavior and real user behavior.

Critical strengths
  • The problem is important, and the modeling direction is reasonable. The cost of user simulation is high. To get stat-sig results, A/B testing with large population is needed to test a recommendation strategy change.

  • Behavior generation is constrained by observed interactions. UGW with uniform initialization, and prior-guided UGW. Decision Accuracy increases from 0.242 for Direct BC to 0.278 for prior-guided UGW, providing preliminary evidence for the alignment procedure.

  • The controlled baseline comparison shows measurable improvements. In Table 4, all simulators receive the same user histories, item metadata, and candidate pages. Compared with DyTA4Rec, PsyBer-Agent improves Decision Accuracy from 0.254 to 0.275 and reduces JSD from 0.335 to 0.286. These results support its relative advantage under the evaluated protocol.

Critical weaknesses
  • Long-horizon behavioral fidelity remains insufficiently validated, and the reported results show substantial degradation. Long-horizon closed-loop simulation is a central motivation, but the supporting evidence is limited: The human–simulator agreement study covers only 10 browsing rounds. In modern recommendation system, ultra-long sequence learning is the central problem, a typical user is usually represented by a sequence of O(10k). The sequence length in this work is limited.

  • Evaluation dataset is limited. The current data is built from 2 scenario: shopping and web browsing, while the modern recommendation system plays a more important role in social media with high frequency scenario like short-form video and news feeds, where catching user's interests are more important. Also, the key claim from the paper like "underlying user dynamics such as fatigue, affect, and shifting preferences", are easier to capture. In [1,2], such dataset is provided so it's interesting to check if the proposed approach can generalize.

[1] Fostiropoulos, Iordanis, et al. "GISTBench: Evaluating LLM User Understanding via Evidence-Based Interest Verification." arXiv preprint arXiv:2603.29112 (2026).

[2] Jiang, Bowen, et al. "PersonaMem-v3: Toward Omni-Platform Personal Intelligence for Holistic User Understanding, Recommendation, and Agentic Tasks." arXiv preprint arXiv:2608.21381 (2026).

Questions for authors
  • With websim, is it possible to build long user history sequence? or is it better to utilize real-user distributed dataset like [1] and [2]?

  • Which scenario is better to study user dynamics such as fatigue, affect, and shifting preferences?

[1] Fostiropoulos, Iordanis, et al. "GISTBench: Evaluating LLM User Understanding via Evidence-Based Interest Verification." arXiv preprint arXiv:2603.29112 (2026).

[2] Jiang, Bowen, et al. "PersonaMem-v3: Toward Omni-Platform Personal Intelligence for Holistic User Understanding, Recommendation, and Agentic Tasks." arXiv preprint arXiv:2608.21381 (2026).

4 comments

Dear Reviewer,

Thank you for your thoughtful feedback. We have carefully addressed the concerns raised in our response. If our clarifications have addressed your concerns, we would be grateful if you would consider updating your evaluation accordingly. Please let us know if any further clarification would be helpful.

Reviewer_kevvOpenReview ↗

Thanks for the update

Thanks for running the evaluation. The results are a bit confusing: when seq len increased from 10 to 100, the decision accuracy drops. This is a bit contradictory to the scaling law - model improves as the user behavior sequence gets longer because the model knows more context about the user

Thank you for pointing this out. In our table, 10, 50, and 100 refer to the number of evaluated steps. The results do not compare predictions for the same target using progressively longer user histories. Longer evaluations include later decisions, so average accuracy may decline even if more history improves prediction for a fixed target. Both PsyBer-Agent and the calibrated baseline show this decline.

Reviewer_kevvOpenReview ↗

Thanks for following up

The results do not compare predictions for the same target using progressively longer user histories

Got it, "same target using progressively longer user histories" is a standard evaluation protocol for recsys research so a table with this will be more persuasive. Given the current status of this manuscript, I tend to keep my original ratings

Reviewer_PHG4OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PsyBer-Agent, a psychology-inspired user simulator for recommender systems. Its Psy-Engine models interaction dynamics through three evolving, model-defined latent states representing cognition, emotional satisfaction, and self-congruence. The framework reconstructs reference state trajectories from logged user interactions through the same Psy-Engine and uses Unbalanced Gromov–Wasserstein alignment to derive supervision for a lightweight action policy, with LLM/VLM fallback when needed. The authors also introduce WebSim and the PsyBer Benchmark for simulation and evaluation across recommendation scenarios and backbones. Experiments report improved behavioral fidelity over the evaluated baselines under a fixed-candidate replay setting, alongside component ablations, robustness analyses, and scalability tests with up to five million queued simulation tasks.

Critical strengths
  1. The paper builds WebSim, an end-to-end closed-loop simulation platform that integrates multimodal asset preparation, offline alignment, and browser-level interactive simulation. It supports evaluation across multiple datasets and recommender backbones, with scalability tests covering million-scale queued workloads under bounded concurrency. The engineering contribution is substantial.
  2. The paper presents a concrete pipeline for converting offline interaction logs into supervision for an online action policy through model-defined reference trajectories and Unbalanced Gromov–Wasserstein alignment, without fine-tuning the backbone LLM.
Critical weaknesses
  1. The use of HR@10 and NDCG@10 to assess human–simulator agreement is insufficiently justified. Both groups receive identical replayed recommendation slates and share the same held-out relevance targets, so under the stated definitions, these metrics appear to evaluate the recommender, not differences between human and simulated behavior. Without clarifying how the decision-maker’s actions affect their computation, the small metric gaps in Table 2 do not provide convincing evidence of behavioral agreement.
  2. The experiments do not sufficiently isolate the contribution of the psychology-inspired state design. Removing Psy-Engine shows that the module is useful, but does not establish whether its specific three-state formulation offers benefits over a generic learned state representation using the same information and training budget. This limits the evidence for the proposed psychological inductive bias.
Questions for authors
  1. How do human and simulated actions affect the computation of HR@10 and NDCG@10 in Table 2, given identical replayed recommendation slates and the same held-out relevance targets?
  2. What evidence distinguishes the benefit of the specific three-state design from that of a simpler state representation using the same input information and training budget?
  3. Given that long-horizon simulation is a central motivation, could the authors discuss the longer-horizon findings and their limitations more explicitly in the main text, particularly what they establish about accumulated behavioral drift?
1 comment

Dear Reviewer,

Thank you for your thoughtful feedback. We have carefully addressed the concerns raised in our response. If our clarifications have addressed your concerns, we would be grateful if you would consider updating your evaluation accordingly. Please let us know if any further clarification would be helpful.

Reviewer_NyjAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PsyBer-Agent, an LLM-based user simulator for recommender systems that replaces purely prompt-driven role-play with an explicit, model-defined latent psychological state.

Critical strengths

The paper provides a practical way to model changes in user behavior over repeated interactions. Its three-dimensional state is easy to inspect, and the alignment procedure uses logged interactions to train a lightweight policy. The ablations show that the state dynamics, alignment, and memory all improve fidelity. The implementation is also a strength: it achieves lower latency and demonstrates execution on large queued workloads.

Critical weaknesses
  1. Reference trajectories and CES-MSE targets are generated by the same Psy-Engine, so latent-state consistency provides internal validation only. Although the behavioral results seem to support the overall pipeline, they do not establish whether its benefits depend on the C/E/S formulation or could be obtained with a comparable nonpsychological state representation. The existing ablations do not resolve this distinction, which kind of make the paper’s central conceptual contribution insufficiently substantiated.
  2. PsyBer-Agent learns from target-domain logs through alignment and policy training, but it is unclear whether the compared simulators receive comparable calibration. Consequently, Table 4 may combine gains from the proposed state design with gains from additional supervised adaptation. Moreover, the only cross-simulator comparison (Table 4) is confined to MM-ML-1M with 10 decision steps per user while the benchmark itself spans seven datasets, no baseline simulator is evaluated beyond this single scenario, so the comparative evidence does not support the paper’s long-horizon, multi-scenario claims.
Questions for authors
  1. Can you provide a control that keeps the identical UGW alignment and fast-policy training but replaces the Psy-Engine C/E/S dynamics with (a) a learned low-dimensional state, or (b) randomly parameterized bounded dynamics of the same dimensionality? This would directly test whether the psychology-inspired structure, rather than calibration plus any structured state, drives the gains. The "w/o Psy-Engine" ablation in Table 6 removes the structured state entirely, which does not answer this question.

  2. In Table 4, what dataset-specific signal (persona construction, few-shot examples, memory initialization) do the baselines receive? If none, how should readers separate PsyBer-Agent's calibration-data advantage from its architectural advantage? Could one baseline be calibrated on the same split for a fairer comparison?

2 comments

Dear Reviewer,

Thank you for your thoughtful feedback. We have carefully addressed the concerns raised in our response. If our clarifications have addressed your concerns, we would be grateful if you would consider updating your evaluation accordingly. Please let us know if any further clarification would be helpful.

Area_Chair_xpvhOpenReview ↗

Gentle Reminder on Rebuttal Discussion

Dear Reviewer,

As the discussion deadline is approaching, I would kindly encourage you to review the authors' rebuttal and share your follow-up comments, particularly on whether your concerns have been adequately addressed.

Your feedback would be greatly appreciated and helpful for the final assessment. Thank you for your time and contribution!

Comments on the paper

2 comments
AuthorsOpenReview ↗

Thank You to the Area Chair and Reviewers

Dear Area Chair and Reviewers,

As this stage of the discussion draws to a close, we would like to thank the Area Chair and all three reviewers for their time and constructive feedback. Your questions helped us improve the paper and explain its contributions and limitations more clearly.

We appreciate Reviewer NyjA’s questions about the psychological state design and the fairness of baseline comparisons, which encouraged us to strengthen the supporting evidence. We thank Reviewer PHG4 for recognizing the engineering contribution and for prompting us to clarify the evaluation and the role of the proposed state representation. We are also grateful to Reviewer kevv for emphasizing longer-term behavior and broader recommendation scenarios, and for continuing to engage with our responses.

We thank the Area Chair for overseeing the discussion and all three reviewers for the attention they have given our work. Their feedback has helped us make the paper more precise and better supported.

Sincerely,
Authors of Submission 24309

Area_Chair_xpvhOpenReview ↗

Reminder: Reviewer–Author Discussion

Thank you to the reviewers who have already engaged in the discussion and provided valuable feedback. As the discussion deadline is approaching, I would kindly encourage the remaining reviewers to review the authors' responses and share their follow-up comments, particularly on whether their concerns have been adequately addressed.

Authors are also encouraged to respond promptly to any further questions. Thank you all for your time and contributions to a constructive discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.