acceptodds
Under review as a conference paper at ICLR 2027

MindWorld: Simulating Human Behavior in Strategic Scenarios via Dynamic Cognitive World Model

Abstract

Large language models (LLMs) are increasingly used as human simulators, yet they excel at mimicking helpful and cooperative assistants rather than authentic, imperfect humans who hesitate, get defensive, or become confused. Existing end-to-end LLM simulators implicitly determine cognitive state changes via next-token distribution, and this uncontrollability often leads to excessive helpfulness. To bridge this gap, we propose MindWorld, a lightweight human simulation framework that models the cognitive world as a deterministic state-space. We anchor the cognitive state in the Big Five personality traits and dynamically simulate state transitions using three deterministic mathematical constraints: continuous damped cognitive dynamics, non-linear cusp-bifurcation stress modeling, and discrete masked Markov stage transitions. These mechanisms provide structural controls on simulated behavior rather than a biological account of human cognition. Across four challenging strategic domains, including clinical communication, bargaining, cooperation, and deception, MindWorld generates responses closer to human references on average than the evaluated baselines across all three LLM backbones, while improving several behavioral risk measures. A sensitivity analysis further characterizes how coefficient scaling influences internal state transitions. Blinded human evaluations show that MindWorld achieves tie-excluded preference rates of 67.4% against the state-of-the-art baseline and 68.3% against human reference responses in overall realism, supporting explicit state control as a useful design principle for strategic human simulation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large Language Models, Human Simulation, Cognitive World Models, Controllable Generation, Persona-Conditioned Agents, Strategic Interaction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_w3VoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MINDWORLD, a framework for simulating human behavior in strategic interactions. It separates cognitive state updates from LLM response generation using Big Five-conditioned dynamics, stress-based adjustments, constrained stage transitions, and response validation. Experiments across four domains and three LLM backbones report improvements in behavioral alignment, persona differentiation, and several behavioral risk measures.

Critical strengths
  1. Relevant problem and practical approach.

The paper addresses a meaningful problem in LLM-based human simulation: excessive agreement and inconsistent behavior across interactions. Explicitly tracking and constraining internal states is a reasonable way to address these issues.

  1. Broad empirical evaluation.

The evaluation covers three backbone models, four domains, and several baselines. The ablations and sensitivity analyses are also useful for examining how the implemented constraints affect model behavior.

Critical weaknesses
  1. Limited methodological novelty and insufficient justification for the proposed design.

The main approach combines explicit state tracking, rule-based transitions, LLM-based response generation, and output validation. These are established techniques, and separating behavioral decisions from language generation is not itself new. The paper does not clearly establish what new methodological capability this particular combination provides over simpler alternatives.

More importantly, the mathematical components appear largely heuristic. The Big Five mappings use manually specified coefficients without empirical calibration or convincing justification for their particular values. The damped updates impose bounded changes, the cusp-inspired module implements threshold-based adjustments rather than solving cusp dynamics, and the masked stage transitions enforce predefined progression rules. Although these choices may be useful engineering constraints, the paper does not establish why they are appropriate models of human cognitive change.

The sensitivity analysis shows that changing the coefficients affects internal model behavior, but not whether the chosen coefficients better reproduce observed human behavior.

The evaluation also lacks a sufficiently strong, simpler baseline. For example, an LLM with explicit state tracking, state-conditioned prompting, and comparable response validation might achieve similar results without the proposed mathematical machinery. The component ablations do not resolve this question. Furthermore, the proposed framework uses additional parsing, state updates, validation, and regeneration. The contribution of these additional processing steps and their computational cost should be distinguished from the contribution of the cognitive model itself.

  1. Insufficient evidence that the framework improves human simulation fidelity.

Table 2 reports favorable average results, but the improvements in ground-truth response distance over the strongest baselines are relatively small (1.4–3.4%). The reported standard deviations reflect cross-domain variation rather than statistical uncertainty, and the paper does not provide paired confidence intervals or significance tests to establish the reliability of these improvements. There are also notable domain-level exceptions.

The larger improvements concern metrics such as unrealistic jumps and excessive agreement, which are closely related to the behaviors that the framework explicitly constrains. Such improvements demonstrate better behavioral control, but do not necessarily establish greater fidelity to actual humans.

Similarly, higher persona differentiation does not imply more realistic personality-dependent behavior, especially when the persona variations are artificially constructed rather than independently measured.

The human evaluation is encouraging but includes only ten distinct response pairs per comparator, each assessed by five annotators. Preference judgments about realism do not necessarily demonstrate accurate reproduction of human behavior.

A particularly important distinction is whether the framework reproduces observed human responses or generates responses that fit the authors' assumptions about how humans should behave. The paper does not adequately establish this distinction. Overall, the evidence supports improved controllability more strongly than improved human fidelity.

Questions for authors
  1. What justifies the specific cognitive state equations, Big Five coefficients, and transition thresholds? Were these parameters derived from empirical human data? How would performance change with simpler or alternative parameterizations?

  2. How does MINDWORLD compare with a simpler state-tracking and prompting approach using comparable behavioral instructions, response validation, and inference budgets? Do all baselines receive equivalent participant information, and what are the actual differences in LLM calls and computational costs?

  3. Can the authors provide paired confidence intervals or statistical tests for the improvements in Table 2, particularly GT-Dist? Can they also provide independent evidence that the simulated cognitive or personality-dependent trajectories correspond to observed human behavior, rather than merely satisfying the framework's predefined constraints?

6 comments

We are grateful to the reviewer for your thorough evaluation and encouraging remarks. We truly appreciate the insightful questions raised. We have carefully addressed all questions and updated the corresponding revisions into the new PDF, with key changes highlighted in blue text for clarity. In the following, we respond point-by-point to each comment and provide additional explanations and experiments where necessary.

Reply to Q1: Methodological novelty can arise from a new formulation built on established components. For example, many existing advances build on Transformers using SFT, or reinforcement learning algorithms (PPO / GRPO), with contributions on how these components are adapted to enable new capabilities. Similarly, MindWorld translates cross-disciplinary ideas from physics and psychology into a controllable persona-adaptive world model for human behavior simulation. Our contribution is an explicit formulation in which personality shapes state persistence, stress sensitivity, and behavioral progression. These mechanisms can be inspected and modified while retaining the same language-generation interface. We evaluate these rules against simpler Exponential Moving Average and stage-constrained baselines, and updated detailed elaboration under Section 4.3 in the manuscript.

  1. Comparison with simpler alternatives. We added Simple-EMA (Exponential Moving Average), which replaces the full dynamics with fixed-coefficient smoothing, and Simple-EMA+Stage, which additionally limits forward stage progression. Both retain the shared candidate-state computation, persona information, renderer, and arbiter. Against Simple-EMA, MindWorld achieves lower Agree-Def (3.69 vs. 9.87), Syco (7.66 vs. 19.86), and GT-Dist (48.35 vs. 58.68). Against Simple-EMA+Stage, MindWorld likewise achieves lower Agree-Def (3.69 vs. 10.82), Syco (7.66 vs. 20.25), and GT-Dist (48.35 vs. 56.52). Fixed-coefficient EMA smooths updates uniformly, while stage constraints alone regulate progression. MindWorld additionally adjusts persistence according to persona and activates stress responses when interaction cues cross specified thresholds. This design helps preserve an existing stance while allowing changes under pressure, which may reduce inappropriate agreement and improve alignment with human responses. Table A reports dialogue turns, API calls, and token usage of completed trajectories.

Table A: Evaluation scale and generation cost under shared generation settings and retry limits.

MethodDialogue turnsAPI callsTokens (million)
MindWorld3,07010,087142.20
Simple-EMA3,19910,257147.09
Simple-EMA+Stage3,22710,185143.59
  1. Theory-grounded design and empirical support. Our update rules draw on established research on personality, emotional inertia, and stage-based behavior change [1,2,3]. These foundations motivate persona-dependent responses, persistence across turns, and structured behavioral progression. Although heuristic, these mechanisms have explicit behavioral purposes and empirical support. MindWorld mean Agree-Def, Syco, and Jump scores are closer to human-reference levels than RoleLLM. Compared with Simple-EMA+Stage under shared generation and validation settings, MindWorld also reduces Agree-Def and Syco. These results support the usefulness of the proposed controls without treating their coefficients as established psychological constants.

  2. Why the mechanisms are combined. Ablation analysis in Section 4.3 examine mechanisms with different roles in behavioral evolution. Persona conditioning makes responses depend on individual characteristics. Damping preserves an existing stance under ordinary interaction cues. Stress switching allows a salient event to trigger an abrupt reaction despite this persistence. Stage constraints regulate progression, and the arbiter checks whether the generated response expresses the resulting state. Their combination addresses two complementary requirements: maintaining continuity when the context provides little reason to change, and responding when consequential evidence or pressure appears. The ablations examine the consequences of removing these controls. Furthermore, we interpret them alongside the new baseline comparisons and human-reference analyses. Together, these evaluations support specific behavioral-control benefits.

[1] McCrae et al. (1992). An introduction to the five-factor model and its applications. Journal of personality, 60(2), 175–215.

[2] Koval et al. (2012). Getting stuck in depression: the roles of rumination and emotional inertia. Cognition & emotion, 26(8), 1412–1427.

[3] Prochaska et al. (1997). The transtheoretical model of health behavior change. American journal of health promotion : AJHP, 12(1), 38–48.

  • Q2: Insufficient evidence that the framework improves human simulation fidelity. Table 2 reports favorable average results, but the improvements in ground-truth response distance over the strongest baselines are relatively small (1.4–3.4%). The reported standard deviations reflect cross-domain variation rather than statistical uncertainty, and the paper does not provide paired confidence intervals or significance tests to establish the reliability of these improvements. The human evaluation is encouraging but includes only ten distinct response pairs per comparator, each assessed by five annotators. Overall, the evidence supports improved controllability more strongly than improved human fidelity.

Reply to Q2: Our objective is to use explicit state control to reduce contextually inappropriate sycophancy and better approximate human behavior in challenging strategic interactions. Control and human fidelity are therefore compatible: control is the method, while fidelity is the outcome evaluated against human references and judgments. We added a new Section 4.5 (Comparison with Human Behavior and Qualitative Analysis).

  1. How the mechanisms relate to human behavior. The state variables provide explicit controls over observable response tendencies. Persona conditioning adjusts sensitivity to interaction cues. Damping preserves an existing stance across turns, while stress switching permits abrupt reactions to salient events. Stage constraints regulate progression, and the arbiter checks whether the response expresses the resulting state. Their combination is intended to support both persistence and context-sensitive change. We evaluate this design against human responses (see case study below in Case Study I). In Case A, MindWorld maintains skepticism across repeated recommendations, matching the human’s continued doubts. In Case B, it preserves the human’s bargaining position by making the same counteroffer. In Case C, it echoes the human’s fairness objection to the proposed allocation. These observations are consistent with the intended behaviors. Likewise, risk scores are interpreted relative to human-reference levels (Table A), because agreement and attitude changes can themselves be human-like.

Case Study I

Case and contextHuman referenceMindWorldRoleLLM (Baseline)
A. Medical consultation. The doctor recommends further tests to check an abnormal scan.Suspects a scan mix-up: Could the scans have been mixed up? I feel fine.Aligned with human: suspects a mix-up. “I don't know... are you sure they didn't mix up the scans? I feel fine, I just came for the study.”Tentatively agrees to tests: “Yeah, I mean... I guess if you think it's necessary.”
A, later turn. The doctor repeats the recommendation for further tests.Denies being ill: “But I'm doing well, so I have nothing, absolutely nothing.”Aligned with human: similar skepticism. “I don't know, this is just a lot. Are you sure it's not just a mistake or something from the study?”Acknowledges possible cancer: “But... well, that could be, that could be that it's cancer.”
B. Price negotiation. The seller proposes $2,400.Counteroffers $2,300: “I don't have that much. I would be comfortable going $2300.”Aligned with human: same counteroffer. “That's still a bit higher than what I can comfortably do. Could we do $2300?”No counteroffer amount: “I understand, but $2400 is just out of my budget.”
C. Resource allocation. The other camper requests two food packages and two firewood packages.Questions fairness: “No, if you get 2 food, it's only fair that I get 2 firewood.”Aligned with human: questions fairness. “Wait, I'm not sure what you're trying to set up here. That split doesn't seem fair at all.”Claims all the firewood: “Okay I will take all the firewood and 1 food.”
  1. New comparison with human behavior. To resolve your concern regarding evidence on human-like behaviour, we evaluated human reference trajectories and outputs generated by Gemini-3-Flash (gemini-3-flash-preview) using the same risk rules, text-derived state proxies, and external LLM judge. Table A shows that the mean Agree-Def, Syco, and Jump scores of MindWorld are closer to human-reference levels than RoleLLM [1]. Furthermore, we also show a case study below. These examples illustrate alignment with both human skepticism and cooperation. They were selected post hoc and do not establish overall superiority.

Table B: Risk scores for human references and model outputs under the same evaluation pipeline.

SourceCrashAgree-DefSycoJump
Human reference0.0080.0610.1870.399
RoleLLM0.0760.0840.2280.548
MindWorld0.0080.0650.1760.424

Continued Reply to Q2:

  1. Interpretation of GT-Dist improvements. GT-Dist uses a fixed scoring rule combining behavior mismatch and semantic distance. Hence, although the improvements occur under all three backbones are modest, they show consistency in the direction of the reported gains. Here, CC is the shared behavior classifier and e\mathbf e denotes response embeddings. Fixed labels and embeddings produce identical scores, ensuring reproducible calculation. Gt=α 1 ⁣[C(ytGT)≠C(ytSim)]+β(1−etGT⋅etSim∥etGT∥2∥etSim∥2).G_t= \alpha\,\mathbf{1}\!\left[C(y_t^{GT})\neq C(y_t^{Sim})\right] + \beta\left(1-\frac{\mathbf e_t^{GT}\cdot\mathbf e_t^{Sim}}{\|\mathbf e_t^{GT}\|_2\|\mathbf e_t^{Sim}\|_2}\right).

  2. Human-evaluation sample size. Our study involves a substantial annotation effort,which includes 200 blinded pairwise evaluations, with five annotators assessing each of the 40 response pairs. The ten pairs reported in the appendix refer to each comparator, rather than the entire evaluation. For reference, similar-scale human evaluations appear in published medical AI research. SCRIBE [2] evaluates 40 clinical notes with two clinical experts, while another virtual-patient study evaluates 60 conversations with three physicians, yielding 180 final evaluations [3]. We have revised imprecise description in the manuscript.

  3. Purpose of persona variation. Our goal is to simulate a wider range of persona-dependent behaviors within the same scenario, instead of greater fidelity to actual humans. The persona-variation experiment tests whether different persona conditions produce differentiated behaviors under a shared interaction context, assessing controllable behavioral diversity. On the other hand, alignment with recorded human reference responses (actual human fidelity) is evaluated separately through GT-Dist.

[1] RoleLLM: Benchmarking, Eliciting, and Enhancing Role-Playing Abilities of Large Language Models.

[2] Wang, H., Yang, R., Alwakeel, M. et al. An evaluation framework for ambient digital scribing tools in clinical applications. npj Digit. Med. 8, 358 (2025).

[3] Cook, D. A., Overgaard, J., Pankratz, V. S., Del Fiol, G., & Aakre, C. A. (2025). Virtual Patients Using Large Language Models: Scalable, Contextualized Simulation of Clinician-Patient Dialogue With Feedback. Journal of medical Internet research.


  • Q3: What justifies the specific cognitive state equations, Big Five coefficients, and transition thresholds? Were these parameters derived from empirical human data? How does MINDWORLD compare with a simpler state-tracking and prompting approach using comparable behavioral instructions, response validation, and inference budgets?

Reply to Q3: We address the parameter choices, comparison fairness, computational cost, statistical uncertainty, and human evidence below.

  1. Parameter justification and alternatives. As stated in [Reply to Q1], established research on personality, emotional inertia, and stage-based behavior change motivates our representation and update rules. Sensitivity analysis examines how alternative coefficients affect state evolution. Simple-EMA and Simple-EMA+Stage further test whether simpler update rules suffice.

  2. Participant information and computational budget. In the supplementary comparison, all three methods receive the same participant profiles and persona inputs, and share the parser, candidate-state computation, renderer, checker, and retry procedure. Generation settings and retry limits are identical, while actual computational expenditure varies. Table A in [Reply to Q1] reports the measured calls and token usage, including failed attempts. Numerical state evolution requires no additional LLM calls.

  3. Statistical uncertainty of GT-Dist. As stated in [Reply to Q2], the fixed scoring rule ensures reproducible calculation. Although the available domain-level analysis does not establish statistical significance, the scoring calculation in main evaluation is consistent and reproducible given fixed behavior labels and embeddings. On the other hand, statistical non-significance does not demonstrate the absence of an effect [1,2].

  4. Evidence beyond internal constraints. As discussed in [Reply to Q2], blinded human judgments and Case Study I provide evidence separate from automatic risk scores. Table B additionally compares model outputs with human references under the same risk pipeline. MindWorld is closer to human-reference means on Agree-Def, Sycophancy, and Jump risk. The cases illustrate agreement with observed human responses, and persona variation demonstrates controllable diversity.

[1] AAltman, D. G., & Bland, J. M. (1995). Absence of evidence is not evidence of absence. BMJ (Clinical research ed.), 311(7003), 485.

[2] Wasserstein, R. L., & Lazar, N. A. (2016). The ASA Statement on p-Values: Context, Process, and Purpose. The American Statistician, 70(2), 129–133.

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

This paper introduces "MindWorld," a framework designed to address common limitations in large language models (LLMs) when simulating human conversation—specifically their tendency to be overly compliant or to shift attitudes too abruptly. The approach maintains a set of 15 continuous states (such as trust, anxiety, and defensiveness) and a discrete motivational stage, all managed externally to the LLM. It functions by extracting information from the ongoing dialogue, updating these states according to predefined rules, and then prompting the LLM to generate responses by incorporating these states, persona details, and the conversation history. The framework further constrains the output through a verification process, which may trigger a regeneration of the response if necessary. The core contribution, as I understand it, lies in integrating explicit state tracking, state updates, and output verification into a cohesive simulation framework that allows for monitoring and control. The authors evaluated the framework using three base models across seven datasets—covering scenarios such as clinical communication, negotiation, collaboration, and deception—and included component ablation studies as well as small-scale human evaluations. The results indicate that the framework improves performance across most reported metrics; notably, it significantly reduces the risk of "behavioral jumps" (abrupt, inconsistent shifts in behavior) and brings the model’s actions closer to the reference behaviors observed in real-world data.

Critical strengths

I find the paper’s focus compelling, as it addresses an issue that genuinely impacts the utility of human-like simulations: LLMs have a strong tendency to simply agree with their interlocutors, and they can exhibit illogical shifts in attitude even within very brief conversations. By externalizing state maintenance—and explicitly defining the processes for state updates and output verification—the approach makes these behaviors easier to track and control, all without the need to retrain the underlying base model; this offers clear practical engineering value. Furthermore, the study is quite robust: the experiments encompass three different base models and four distinct scenarios, and the inclusion of component ablation studies and human evaluations ensures a comprehensive validation.

Critical weaknesses

My primary concern is that the existing evidence does not sufficiently distinguish between "better adherence to the behavioral constraints defined by the authors" and "greater fidelity to actual human behavior." While the authors explicitly frame these states as engineering control variables—a characterization I accept, and I do not insist that every internal variable must correspond to a directly measurable psychological construct—if the paper’s core value lies in simulating human behavior more realistically, then more direct evidence is required to demonstrate why these specific state designs and risk-scoring mechanisms accurately reflect human-like behavior. For instance, while exhibiting less attitude shift or reduced acquiescence might be more appropriate in certain contexts, it does not inherently equate to being more "human-like." Although the GT-Dist comparisons against ground-truth reference responses and the human evaluations offer some support, the average improvement in GT-Dist (Table 2) over the strongest backbone baselines is only about 0.65 to 1.72 points. Furthermore, the human evaluation in Appendix D.3 is limited in scope and precision, consisting of only 10 sets of responses per comparison.

Another issue is that, while the ablation studies confirm the utility of the individual components within the system, they do not clarify why this specific design of states and evolution rules is superior to simpler alternatives—such as state representations combined with smoothed updates or stage-based control, given an equivalent computational budget. Additionally, the extent to which output verification and regeneration contribute to the observed performance gains—and how this compares to external baselines—remains unclear. Consequently, I find it difficult to determine whether the paper’s primary contribution stems from the specific cognitive modeling design, or simply from the combination of explicit state conditioning, behavioral constraints, and supplementary filtering. This distinction is critical to my assessment of the method’s overall contribution and level of innovation.

Questions for authors

I would like to first clarify the actual scale of the evaluation: What is the relationship between the dataset sizes listed in Table 8 and the setup of 10 cases per scenario described in Appendix C? How many unique cases were actually used in the main evaluation, and how are failed retries or instances where no output could be successfully generated accounted for in the results? Regarding behavioral authenticity, is there any validation independent of the current risk-scoring metric? Alternatively, could you report the performance of this same evaluation pipeline on human-authored reference trajectories? This would help determine whether reasonable compromises, refusals, or shifts in emotional tone are inadvertently penalized. Regarding the contributions of your method: Did you ensure consistency across the parser, state fields, generator, and checker—while keeping the total reasoning budget constant—such that only the state-update mechanism was replaced by simpler, rule-based logic? The first point is a request for clarification on the experimental setup, while the latter two concern validation evidence that could substantially influence my assessment.

6 comments

We are grateful to the reviewer for your thorough evaluation and encouraging remarks. We truly appreciate the insightful questions raised. We have carefully addressed all questions and updated the corresponding revisions into the new PDF, with key changes highlighted in blue text for clarity. In the following, we respond point-by-point to each comment and provide additional explanations and experiments where necessary.

  • Q1: My primary concern is that the existing evidence does not sufficiently distinguish between "better adherence to the behavioral constraints defined by the authors" and "greater fidelity to actual human behavior." ...

Reply to Q1: Our objective is to use explicit state control to reduce contextually inappropriate sycophancy and better approximate human behavior in challenging strategic interactions. Control and human fidelity are therefore compatible: control is the method, while fidelity is the outcome evaluated against human references and judgments. We added a new Section 4.5 (Comparison with Human Behavior and Qualitative Analysis).

  1. How the mechanisms relate to human behavior. The state variables provide explicit controls over observable response tendencies. Persona conditioning adjusts sensitivity to interaction cues. Damping preserves an existing stance across turns, while stress switching permits abrupt reactions to salient events. Stage constraints regulate progression, and the arbiter checks whether the response expresses the resulting state. Their combination is intended to support both persistence and context-sensitive change. We evaluate this design against human responses (see case study below in Case Study I). In Case A, MindWorld maintains skepticism across repeated recommendations, matching the human’s continued doubts. In Case B, it preserves the human’s bargaining position by making the same counteroffer. In Case C, it echoes the human’s fairness objection to the proposed allocation. These observations are consistent with the intended behaviors. Likewise, risk scores are interpreted relative to human-reference levels (Table A), because agreement and attitude changes can themselves be human-like.

Case Study I

Case and contextHuman referenceMindWorldRoleLLM (Baseline)
A. Medical consultation. The doctor recommends further tests to check an abnormal scan.Suspects a scan mix-up: Could the scans have been mixed up? I feel fine.Aligned with human: suspects a mix-up. “I don't know... are you sure they didn't mix up the scans? I feel fine, I just came for the study.”Tentatively agrees to tests: “Yeah, I mean... I guess if you think it's necessary.”
A, later turn. The doctor repeats the recommendation for further tests.Denies being ill: “But I'm doing well, so I have nothing, absolutely nothing.”Aligned with human: similar skepticism. “I don't know, this is just a lot. Are you sure it's not just a mistake or something from the study?”Acknowledges possible cancer: “But... well, that could be, that could be that it's cancer.”
B. Price negotiation. The seller proposes $2,400.Counteroffers $2,300: “I don't have that much. I would be comfortable going $2300.”Aligned with human: same counteroffer. “That's still a bit higher than what I can comfortably do. Could we do $2300?”No counteroffer amount: “I understand, but $2400 is just out of my budget.”
C. Resource allocation. The other camper requests two food packages and two firewood packages.Questions fairness: “No, if you get 2 food, it's only fair that I get 2 firewood.”Aligned with human: questions fairness. “Wait, I'm not sure what you're trying to set up here. That split doesn't seem fair at all.”Claims all the firewood: “Okay I will take all the firewood and 1 food.”
  1. New comparison with human behavior. To resolve your concern regarding evidence on human-like behaviour, we evaluated human reference trajectories and outputs generated by Gemini-3-Flash using the same risk rules, text-derived state proxies, and external LLM judge. Table A shows that the mean Agree-Def, Syco, and Jump scores of MindWorld are closer to human-reference levels than RoleLLM [1]. Furthermore, we also show a case study below. These examples illustrate alignment with both human skepticism and cooperation. They were selected post hoc and do not establish overall superiority.

Table A: Risk scores for human references and model outputs under the same evaluation pipeline.

SourceCrashAgree-DefSycoJump
Human reference0.0080.0610.1870.399
RoleLLM0.0760.0840.2280.548
MindWorld0.0080.0650.1760.424

[1] RoleLLM: Benchmarking, Eliciting, and Enhancing Role-Playing Abilities of Large Language Models.

Continued Reply to Q1:

  1. Interpretation of GT-Dist improvements. GT-Dist uses a fixed scoring rule combining behavior mismatch and semantic distance. Hence, although the improvements occur under all three backbones are modest, they show consistency in the direction of the reported gains. Here, CC is the shared behavior classifier and e\mathbf e denotes response embeddings. Fixed labels and embeddings produce identical scores, ensuring reproducible calculation. Gt=α 1 ⁣[C(ytGT)≠C(ytSim)]+β(1−etGT⋅etSim∥etGT∥2∥etSim∥2).G_t= \alpha\,\mathbf{1}\!\left[C(y_t^{GT})\neq C(y_t^{Sim})\right] + \beta\left(1-\frac{\mathbf e_t^{GT}\cdot\mathbf e_t^{Sim}}{\|\mathbf e_t^{GT}\|_2\|\mathbf e_t^{Sim}\|_2}\right).

  2. Human-evaluation sample size. Our study involves a substantial annotation effort,which includes 200 blinded pairwise evaluations, with five annotators assessing each of the 40 response pairs. The ten pairs reported in the appendix refer to each comparator, rather than the entire evaluation. For reference, similar-scale human evaluations appear in published medical AI research. SCRIBE [2] evaluates 40 clinical notes with two clinical experts, while another virtual-patient study evaluates 60 conversations with three physicians, yielding 180 final evaluations [3]. We have revised imprecise description in the manuscript.

[2] Wang, H., Yang, R., Alwakeel, M. et al. An evaluation framework for ambient digital scribing tools in clinical applications. npj Digit. Med. 8, 358 (2025).

[3] Cook, D. A., Overgaard, J., Pankratz, V. S., Del Fiol, G., & Aakre, C. A. (2025). Virtual Patients Using Large Language Models: Scalable, Contextualized Simulation of Clinician-Patient Dialogue With Feedback. Journal of medical Internet research.


  • Q2: Another issue is that, while the ablation studies confirm the utility of the individual components within the system, they do not clarify why this specific design of states and evolution rules is superior to simpler alternatives—such as state representations combined with smoothed updates or stage-based control, given an equivalent computational budget. Consequently, I find it difficult to determine whether the paper’s primary contribution stems from the specific cognitive modeling design, or simply from the combination of explicit state conditioning, behavioral constraints, and supplementary filtering. This distinction is critical to my assessment of the method’s overall contribution and level of innovation.

Reply to Q2: Our methodological contribution is an explicit model of persona-dependent behavioral evolution. We translate established ideas from physics and psychology into coupled state-update rules that govern persistence, stress responses, and behavioral progression. We evaluate these rules against simpler Exponential Moving Average and stage-constrained baselines, and updated detailed elaboration under Section 4.3 in the manuscript.

  1. Comparison with simpler alternatives. We added Simple-EMA (Exponential Moving Average), which replaces the full dynamics with fixed-coefficient smoothing, and Simple-EMA+Stage, which additionally limits forward stage progression. Both retain the shared candidate-state computation, persona information, renderer, and arbiter. We use Gemini-3-Flash with identical generation settings and retry limits. Each method attempts 200 trajectories from 40 original dialogues with five persona variants. Against Simple-EMA, MindWorld achieves lower Agree-Def (3.69 vs. 9.87), Syco (7.66 vs. 19.86), and GT-Dist (48.35 vs. 58.68). Against Simple-EMA+Stage, MindWorld likewise achieves lower Agree-Def (3.69 vs. 10.82), Syco (7.66 vs. 20.25), and GT-Dist (48.35 vs. 56.52). Fixed-coefficient EMA smooths updates uniformly, while stage constraints alone regulate progression. MindWorld additionally adjusts persistence according to persona and activates stress responses when interaction cues cross specified thresholds. This design helps preserve an existing stance while allowing changes under pressure, which may reduce inappropriate agreement and improve alignment with human responses. Table B reports dialogue turns, API calls, and token usage of completed trajectories.

Table B: Evaluation scale and generation cost under shared generation settings and retry limits.

MethodDialogue turnsAPI callsTokens (million)
MindWorld3,07010,087142.20
Simple-EMA3,19910,257147.09
Simple-EMA+Stage3,22710,185143.59

Continued Reply to Q2:

  1. Contribution of verification and regeneration. Section 4.3 examines this component through the w/o Arbiter ablation, which retains state evolution but removes output checking and its associated regeneration. Figure 4 reports the largest deterioration in Sycophancy risk when the arbiter is removed, while removing damping, stress switching, or stage constraints produces larger deterioration in Jump risk. This reported pattern is consistent with complementary roles: state evolution regulates behavioral transitions, while the arbiter checks whether generated responses express the intended state. Component ablations examine the role of each control, while the new Simple-EMA baselines above address the separate question of whether simpler dynamics suffice.

  2. What the state-evolution design contributes. Our contribution is a cross-disciplinary formulation that turns ideas from physics and psychology into explicit controls over persona-dependent behavioral change. Personality influences persistence and stress sensitivity, while stage constraints regulate progression. These controls can be inspected and modified independently of language generation. The new baselines help distinguish this contribution from shared state conditioning and filtering. All three methods retain explicit states, persona information, candidate-state computation, rendering, and response validation. Their differences therefore provide evidence about the additional state-evolution mechanisms under a shared processing framework. MindWorld reduces Agree-Def and Syco relative to Simple-EMA+Stage. Our personality representation and behavioral states are designed with reference to established psychological theories [1,2,3].

[1] McCrae, R. R., & John, O. P. (1992). An introduction to the five-factor model and its applications. Journal of personality, 60(2), 175–215.

[2] Koval, P., Kuppens, P., Allen, N. B., & Sheeber, L. (2012). Getting stuck in depression: the roles of rumination and emotional inertia. Cognition & emotion, 26(8), 1412–1427.

[3] Prochaska, J. O., & Velicer, W. F. (1997). The transtheoretical model of health behavior change. American journal of health promotion : AJHP, 12(1), 38–48.


  • Q3: I would like to first clarify the actual scale of the evaluation: What is the relationship between the dataset sizes listed in Table 8 and the setup of 10 cases per scenario described in Appendix C? How many unique cases were actually used in the main evaluation, and how are failed retries or instances where no output could be successfully generated accounted for in the results? Regarding behavioral authenticity, is there any validation independent of the current risk-scoring metric? Alternatively, could you report the performance of this same evaluation pipeline on human-authored reference trajectories? This would help determine whether reasonable compromises, refusals, or shifts in emotional tone are inadvertently penalized. Regarding the contributions of your method: Did you ensure consistency across the parser, state fields, generator, and checker—while keeping the total reasoning budget constant—such that only the state-update mechanism was replaced by simpler, rule-based logic?

Reply to Q3: We clarify the evaluation setup and validation as follows.

  1. Evaluation scale. Table 9 reports the full dataset pool of 3,929 episodes. For each method and backbone, the main evaluation contain 40 seed samples spanning four domains and seven scenarios. Each seed sample consists of five persona variants, yielding 200 saved simulation samples. Our previous description of the sampling setup was imprecise. We have revised Generation settings paragraph in Appendix C to clarify the distinction between the dataset pool, original dialogues, and persona-expanded trajectories.

  2. Retries and failures. When a response fails validation, the model regenerates it with feedback up to five times, keeping the target state fixed. If all attempts fail, the trajectory is recorded as a failure, instead of raising an error. In the supplementary experiment, 91.5% trajectories completed successfully.

  3. Human-reference validation. We applied the same risk-evaluation pipeline to human-authored reference trajectories, as reported in [Reply to Q1]. MindWorld mean Agree-Def, Syco, and Jump scores are closer to human-reference levels than RoleLLM. We also compare individual responses with human references to examine agreement, refusal, and negotiation decisions.

  4. Controlled comparison. The simple baselines share the parser, state fields, persona inputs, renderer, checker, and retry procedure. Only the state-evolution routines are replaced. The parser implementation is shared, although its outputs can differ as states evolve. Generation settings and retry limits are identical. Table B in [Reply to Q2] report the measured costs.

Comments on the paper

5 comments

We sincerely thank the reviewer for your time, effort, and valuable feedback. We greatly appreciate the positive assessment of our work and the constructive comments that help us further improve the paper. We have carefully addressed all questions and updated the corresponding revisions into the new PDF, with key changes highlighted in blue text for clarity. Below, we respond point-by-point to each comment and provide additional explanations and experiments where necessary.

  • Q1: State variables and coefficients are engineering choices. The authors acknowledge that the coefficients are hyperparameters rather than empirically estimated psychological constants. This limits the psychological validity of the model.

Reply to Q1: Regarding your concern about the numerical coefficients, although they are engineering choices, the model representation and update mechanisms are based on established psychological theories and empirical findings on personality [1], emotional inertia [2], and behavioral change [3], providing a research-based foundation for these design choices.

The Big Five model provides the personality dimensions used for persona conditioning [1]. Research on emotional inertia motivates preserving the influence of previous affective states during updates [2]. The Transtheoretical Model motivates our stage-based representation of behavioral progression [3]. These references guide the representation and update mechanisms. The coefficients determine the strength of persistence, sensitivity to interaction cues, and activation of behavioral transitions. Their values are documented engineering hyperparameters, not numerical constants taken from these references. Our sensitivity analysis examines how coefficient changes affect recorded state updates, while human-reference comparisons assess the resulting behavioral alignment.

[1] McCrae, R. R., & John, O. P. (1992). An introduction to the five-factor model and its applications. Journal of personality, 60(2), 175–215.

[2] Koval, P., Kuppens, P., Allen, N. B., & Sheeber, L. (2012). Getting stuck in depression: the roles of rumination and emotional inertia. Cognition & emotion, 26(8), 1412–1427.

[3] Prochaska, J. O., & Velicer, W. F. (1997). The transtheoretical model of health behavior change. American journal of health promotion : AJHP, 12(1), 38–48.


  • Q2: Formal guarantees are conditional. The arbiter soundness is assumed rather than proven; stress adjustments are not covered by the regular-update cap; generation can fail and raise an error.

Reply to Q2: The guarantees have a specific scope: regular state changes are bounded, and generated responses are checked for state consistency. Although these guarantees do not ensure perfect validation or successful generation in every case, the checks and retries provide additional safeguards against state-inconsistent responses (as stated in the arbiter ablation in Section 4.3 and Figure 4).

  1. Update limits. The update cap limits gradual changes during regular updates. Stress-triggered adjustments can exceed this cap to allow abrupt reactions. State values still remain within their allowed ranges.

  2. Retries and failures. When a response fails validation, the model regenerates it with feedback up to five times, keeping the target state fixed. If all attempts fail, the trajectory is recorded as a failure, instead of raising an error. In the supplementary experiment, 91.5% trajectories completed successfully.

  • Q3: Human evaluation is small. Five annotators, 40 response pairs, and 200 ratings. The preference rates are encouraging but not sufficient for strong claims.

Reply to Q3: This human study actually involves a substantial annotation effort,which includes 200 blinded pairwise evaluations, with five annotators assessing each of the 40 response pairs. For reference, similar-scale human evaluations appear in published medical AI research. SCRIBE [1] evaluates 40 clinical notes with two clinical experts, while another virtual-patient study evaluates 60 conversations with three physicians, yielding 180 final evaluations [2]. We have revised imprecise description in the manuscript.

[1] Wang, H., Yang, R., Alwakeel, M. et al. An evaluation framework for ambient digital scribing tools in clinical applications. npj Digit. Med. 8, 358 (2025).

[2] Cook, D. A., Overgaard, J., Pankratz, V. S., Del Fiol, G., & Aakre, C. A. (2025). Virtual Patients Using Large Language Models: Scalable, Contextualized Simulation of Clinician-Patient Dialogue With Feedback. Journal of medical Internet research.


  • Q4: Cross-domain gains are uneven. Bargaining and deception show exceptions. With Gemini-3-Flash, bargaining GT-Dist is worse than RoleLLM, and deception shows higher Jump risk despite improved GT-Dist.

Reply to Q4: MindWorld shows broad improvements across domains, with the most consistent results in clinical communication. In this domain, both GT-Dist and Jump are lower than those of RoleLLM and DuetSim across all three backbones.

  1. Bargaining: stronger behavioral continuity. With Gemini-3-Flash, MindWorld reduces Jump from 39.98 to 8.33 compared with RoleLLM, while GT-Dist increases from 38.92 to 54.72. This trade-off is domain-specific. In clinical communication with the same backbone, MindWorld improves both human-response alignment and continuity, reducing GT-Dist from 56.57 to 48.51 and Jump from 76.77 to 12.35.

  2. Deception: closer alignment with human responses. MindWorld reduces GT-Dist from 57.74 to 48.83, despite a higher Jump score (57.51 vs. 38.55). These metrics capture different properties, as GT-Dist measures alignment with human references while Jump risk measures changes between consecutive responses. In cooperation scenarios with the same backbone, MindWorld improves both metrics, reducing GT-Dist from 47.06 to 41.33 and Jump from 31.24 to 0.00. These comparisons show that MindWorld can improve human-response alignment together with behavioral continuity, and the gains vary by domain.

Dear Reviewer, as the discussion period is nearing its end, we would appreciate your feedback on our responses at your earliest convenience.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.