PsyWorld: Profile-Conditioned World Modeling for Human Behavior Prediction
Summary
Emory University researchers introduce PsyWorld, a profile-conditioned state world model that uses observed environment–behavior state transitions to learn transition-informed representations for predicting human behavior, achieving 0.904 Top-1 accuracy on the VEET behavioral simulator with participant-disjoint generalization.
View Cached Full Text
Cached at: 11/20/26, 09:14 AM
# PsyWorld: Profile-Conditioned World Modeling for Human Behavior Prediction Source: [https://arxiv.org/html/2611.24999](https://arxiv.org/html/2611.24999) Conference:KDD 2027 AI for Sciences Track; August 2027; San Jose, CA, USACCS:Computing methodologies Modeling and simulationCCS:Applied computing PsychologyZiyuan Qinemail:[ziyuan\.qin@emory\.edu](mailto:[email protected])Affiliation:Department of Computer Science,Emory University,Atlanta,Georgia,USAYang Qiaoemail:[yang\.qiao@emory\.edu](mailto:[email protected])Affiliation:Department of Computer Science,Emory University,Atlanta,Georgia,USA,Summer B\. Frandsenemail:[summer\.frandsen@emory\.edu](mailto:[email protected])Affiliation:Department of Psychiatry and Behavioral Sciences,Emory University,Atlanta,Georgia,USA,Michael T\. Treadwayemail:[mtreadway@emory\.edu](mailto:[email protected])Affiliation:Department of Psychiatry and Behavioral Sciences,Emory University,Atlanta,Georgia,USAandLiang Zhaoemail:[liang\.zhao@emory\.edu](mailto:[email protected])Affiliation:Department of Computer Science,Emory University,Atlanta,Georgia,USA 2027 ###### Abstract\. Understanding how stable psychometric traits and time\-varying states jointly inform behavior is an important problem in computational behavioral science\. Existing personalized agents commonly condition actions directly on a static profile, leaving realized state transitions unused as an explicit source of supervision\. We introduce a profile\-conditioned state world model that uses observed environment–behavior state transitions to learn transition\-informed representations for behavioral prediction\. During training, a posterior encoder quantizes realized transitions, while a prior predicts the resulting codes from the psychometric profile and pre\-transition context\. The prior code and predicted next state then condition macro\-level behavior and simulator\-grounded execution\. Thus, the world model provides structured transition supervision instead of treating personality as a direct action descriptor\. We evaluate the framework in VEET, a controlled behavioral simulator with human trajectories, and in AgentSociety as a structurally different synthetic environment\. A participant\-disjoint split of 240 training, 25 validation, and 48 test participants evaluates generalization to unseen people\. On VEET, the model reaches 0\.904 Top\-1 accuracy \(95% CI: \[0\.895, 0\.913\]\) and 0\.0033 Macro KL\. Ablations associate continuous next\-state modeling with most of the Top\-1 gain, while discrete codes and prior alignment mainly improve Top\-3 and distributional alignment\. These results establish transition\-informed world\-model supervision as useful for subject\-generalized behavior prediction, without implying causal psychological mechanisms or calibrated long\-horizon state dynamics\. ###### Keywords: human behavior modeling, psychometrics, world models, personalized agents, out\-of\-distribution generalization, computational behavioral science ## 1\.Introduction Human behavior reflects both stable individual differences and states that vary within a task\. Psychometric profiles describe between\-person variation in stress, motivation, and related traits, whereas recorded task, environmental, and behavioral variables describe the context in which those tendencies are expressed\. Dense longitudinal evidence shows that trait\- and state\-level motivation can exert distinct and interacting effects on effort\-based decisions\([Hewitt et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib1)\)\. For computational behavioral simulation, this distinction suggests that realized state transitions can provide supervision beyond a static profile\-to\-action mapping\. Existing personality\-conditioned agents only partially use this structure\. LLM\-based social simulators combine persona descriptions with needs, memory, and planning\([Piao et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib2);[Wang et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib3)\), while consistency\-oriented training reduces persona drift across interactions\([Abdulhai et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib4)\)\. These systems can produce locally plausible behavior, but profiles are commonly injected as static prompts, policy inputs, or reward descriptors\. A direct profile\-to\-action model can therefore capture participant\-level associations without representing within\-person changes in task, environmental, and behavioral state\. The central challenge is to exploit realized state changes during training without using future information at inference\. Realized environment–behavior state changes provide transition\-level supervision associated with subsequent behavior, but the future state is unavailable before an agent acts\. Under participant\-disjoint evaluation, the model must also transfer this structure to profiles and trajectories never observed during training, rather than memorize participant\-specific action frequencies\. We therefore ask whether explicit state\-transition supervision can improve behavioral prediction for previously unseen participants\. World models provide a natural inductive bias for this question because they learn state evolution rather than only action association\. Latent world models have enabled agents to learn compact transition structure and reason over predicted futures\([Hafner et al\., 2020](https://arxiv.org/html/2611.24999#bib.bib8);[Hafner et al\., 2022](https://arxiv.org/html/2611.24999#bib.bib9);[Hafner et al\., 2023](https://arxiv.org/html/2611.24999#bib.bib10)\)\. Standard world models, however, are designed primarily for task control and do not explicitly separate stable psychometric profiles from the time\-varying environment–behavior state\. Our formulation makes that distinction explicit and uses next\-state prediction as structured supervision for downstream behavior\. We introduce a*profile\-conditioned state world model*\. For each transition, the model separates the stable psychometric profile from the current environment–behavior state and macro behavior\. A training\-time posterior observes the realized state change and assigns it to a vector\-quantized transition code\. A rollout\-time prior predicts that code using only the profile and pre\-transition context, making the transition\-informed representation available without future leakage\. The code and predicted next state condition macro\-level behavior, which is realized through a simulator\-specific micro controller\. The world model is therefore central to the profile–state–behavior factorization, while the evaluated outcome is behavior prediction rather than independent calibration of long\-horizon state dynamics\. We evaluate the formulation in two complementary environments\. VEET provides questionnaire profiles and behavioral trajectories from human participants for the primary unseen\-participant evaluation\. AgentSociety provides a structurally different synthetic LLM\-based environment for testing whether the same factorization can be operationalized across state spaces and action interfaces\. We use a strictly participant\-disjoint split of 240 training, 25 validation, and 48 test participants, so every reported VEET test trajectory comes from a previously unseen person\. Ablations separate the effects of continuous next\-state prediction, vector quantization, and rollout\-time prior alignment; a code–trait analysis bounds the interpretation of the learned representation\. Our contributions are: - •We formulate subject\-generalized human behavior prediction by separating stable psychometric profiles from time\-varying environment–behavior states and using realized transitions as structured supervision\. - •We develop a profile\-conditioned state world model whose posterior extracts vector\-quantized transition codes from observed state changes and whose prior predicts them from pre\-transition context for leakage\-free rollout\. - •On a participant\-disjoint VEET split, we show the predictive utility of transition supervision and characterize the distinct roles of continuous state prediction, discrete codes, and prior alignment; AgentSociety supplies a synthetic operationalization rather than independent psychological validation\. ## 2\.Related Work #### Trait–state modeling and behavioral simulation\. Psychological theories distinguish relatively stable between\-person traits from within\-person states that fluctuate across time and context\. Longitudinal evidence further shows that trait and state motivation can contribute independently and interactively to effort\-based decisions\([Hewitt et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib1)\)\. Simulation platforms provide a controlled setting in which such dynamics can be observed and modeled: Unity\-based experimental frameworks reproduce established behavioral effects\([Brookes et al\., 2019](https://arxiv.org/html/2611.24999#bib.bib5);[Kozlov and Johansen, 2010](https://arxiv.org/html/2611.24999#bib.bib6)\), while interview\-conditioned generative agents can reproduce several empirical patterns in human responses\([Park et al\., 2026](https://arxiv.org/html/2611.24999#bib.bib7)\)\. Our focus is not to assign a fixed psychological type to an agent, but to use observed environmental and behavioral state changes as structured supervision for profile\-conditioned behavior prediction\. #### Personality\-conditioned agents\. Personality\-aware agents commonly inject profile information through prompts, memory, needs, value systems, or policy conditioning\. AgentSociety combines needs, affect, cognition, and memory in large\-scale social simulation\([Piao et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib2)\), whereas D2A maps personality profiles to desire\-driven values for daily activity generation\([Wang et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib3)\)\. Other work has measured how Big Five traits alter LLM\-agent decisions and performance\([Ren and Xu, 2025](https://arxiv.org/html/2611.24999#bib.bib11);[Duan et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib12)\), or used consistency\-based optimization to reduce persona drift across interactions\([Abdulhai et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib4)\)\. These methods establish the value of explicit profiles, but primarily use them as static conditions on behavior\. We instead separate the profile from the time\-varying environment–behavior state and model how their interaction informs subsequent behavior\. #### World models and latent transition learning\. Latent world models learn compact predictive representations of sequential environment dynamics and use them for imagination or control\. Dreamer introduced policy learning in a recurrent latent state space\([Hafner et al\., 2020](https://arxiv.org/html/2611.24999#bib.bib8)\), while discrete and more general successors improved representation learning and cross\-domain robustness\([Hafner et al\., 2022](https://arxiv.org/html/2611.24999#bib.bib9);[Hafner et al\., 2023](https://arxiv.org/html/2611.24999#bib.bib10)\)\. We adopt the predictive principle of world modeling for a different scientific target: learning how questionnaire\-derived traits condition environment–behavior state changes\. Our transition\-informed posterior uses realized future changes to discover discrete modes, and a prior predicts those modes from pre\-transition context for rollout\. The codes therefore represent transition\-level predictive structure rather than static personality categories\. ## 3\.Method ### 3\.1\.Problem Formulation: Transition\-Informed Behavioral Prediction For participantii, the stable questionnaire\-derived profile is𝐪i∈ℝdq\\mathbf\{q\}\_\{i\}\\in\\mathbb\{R\}^\{d\_\{q\}\}\. Each behavioral trajectory is represented by transitions \(𝐬t,𝐪i,at,𝐬t\+1,at\+1\),\(\\mathbf\{s\}\_\{t\},\\mathbf\{q\}\_\{i\},a\_\{t\},\\mathbf\{s\}\_\{t\+1\},a\_\{t\+1\}\),where𝐬t∈ℝ49\\mathbf\{s\}\_\{t\}\\in\\mathbb\{R\}^\{49\}is the recorded environment–behavior state andata\_\{t\}is the current macro behavior\. Our primary empirical objective is p\(at\+1∣𝐪i,𝐬t,at\)p\(a\_\{t\+1\}\\mid\\mathbf\{q\}\_\{i\},\\mathbf\{s\}\_\{t\},a\_\{t\}\)for participants whose profiles and trajectories are absent from training\. Rather than predict the next macro behavior only from current context, we jointly model p\(𝐬t\+1,at\+1∣𝐪i,𝐬t,at\)\.p\(\\mathbf\{s\}\_\{t\+1\},a\_\{t\+1\}\\mid\\mathbf\{q\}\_\{i\},\\mathbf\{s\}\_\{t\},a\_\{t\}\)\.The realized next state supplies structured transition supervision during training\. At inference, a predicted state and prior transition code condition behavior without access to𝐬t\+1\\mathbf\{s\}\_\{t\+1\}\. This factorization treats the latent codes as predictive summaries of profile\-conditioned state transitions, not as causal psychological or clinical constructs\. Figure 1\.A stable psychometric profile conditions latent transition codes together with the current environmental and behavioral state and macro action\. The predicted next\-state representation then informs subsequent behavior prediction\. Dashed components use the realized future only during training; solid paths are available at rollout\.A profile, current environment\-behavior state, and current macro behavior enter a context encoder\. A training\-only posterior observes the realized state change and supplies a vector\-quantized code, while a prior predicts a rollout code from pre\-transition context\. The code conditions a next\-state decoder and a gated macro\-behavior decoder\. ### 3\.2\.Overview The profile\-conditioned state world model first encodes the stable profile, current environment–behavior state, and macro behavior\. A training\-time posterior uses the realized future state change to identify a discrete transition code, while a prior predicts that code from information available before the transition\. Conditioned on the code, the world model predicts the next state\. The transition code and predicted next state feed a macro behavior decoder, and a simulator\-specific execution policy realizes the macro behavior as primitive actions\. The code is therefore a transition\-level predictive variable rather than a participant identity or personality label\. ### 3\.3\.Profile\-Conditioned State World Model We represent stable and time\-varying information separately\. The psychometric profile𝐪i\\mathbf\{q\}\_\{i\}is fixed for participantii, whereas𝐬t\\mathbf\{s\}\_\{t\}records the time\-varying task, environmental, and behavioral state\. The current macro behaviorata\_\{t\}specifies the behavioral context under which the transition occurs\. For AgentSociety, 𝒜AS=\{work,consume,rest,leisure,move\}\.\\mathcal\{A\}\_\{\\mathrm\{AS\}\}=\\\{\\texttt\{work\},\\texttt\{consume\},\\texttt\{rest\},\\texttt\{leisure\},\\texttt\{move\}\\\}\. The context encoder fuses these variables: 𝐡t=fθctx\(𝐪i,𝐬t,at\)\.\\mathbf\{h\}\_\{t\}=f\_\{\\theta\}^\{\\mathrm\{ctx\}\}\(\\mathbf\{q\}\_\{i\},\\mathbf\{s\}\_\{t\},a\_\{t\}\)\.Given a latent transition code𝐳t\\mathbf\{z\}\_\{t\}, the world\-model decoder predicts the next environment–behavior state: 𝐬^t\+1=fθwm\(𝐡t,𝐳t\)\.\\hat\{\\mathbf\{s\}\}\_\{t\+1\}=f\_\{\\theta\}^\{\\mathrm\{wm\}\}\(\\mathbf\{h\}\_\{t\},\\mathbf\{z\}\_\{t\}\)\.This decoder supplies structured transition supervision and makes the profile–state–behavior factorization explicit instead of using the profile only as a direct action descriptor\. ### 3\.4\.Transition\-Informed Latent Representation The latent code must be informative about a realized transition yet available before that transition occurs\. A code learned only from current context can collapse into a static subject embedding or a shortcut for the current state\. A code inferred from the realized future is more transition\-specific, but cannot be used directly at inference\. We address this tension with a prior–posterior scheme\. The rollout prior predicts pθ\(kt∣𝐡t\)=Cat\(softmax\(Wz𝐡t\)\)\.p\_\{\\theta\}\(k\_\{t\}\\mid\\mathbf\{h\}\_\{t\}\)=\\mathrm\{Cat\}\\\!\\left\(\\mathrm\{softmax\}\(W\_\{z\}\\mathbf\{h\}\_\{t\}\)\\right\)\.During training and rollout, we use the hard assignmentktprior=argmaxkpθ\(k∣𝐡t\)k\_\{t\}^\{\\mathrm\{prior\}\}=\\arg\\max\_\{k\}p\_\{\\theta\}\(k\\mid\\mathbf\{h\}\_\{t\}\)and𝐳tprior=𝐜ktprior\\mathbf\{z\}\_\{t\}^\{\\mathrm\{prior\}\}=\\mathbf\{c\}\_\{k\_\{t\}^\{\\mathrm\{prior\}\}\}\. During training, the realized state change Δ𝐬t=𝐬t\+1−𝐬t\\Delta\\mathbf\{s\}\_\{t\}=\\mathbf\{s\}\_\{t\+1\}\-\\mathbf\{s\}\_\{t\}provides privileged transition supervision\. The posterior encoder maps the pre\-transition context and realized change into a continuous latent vector: 𝐳~t=gϕ\(𝐡t,Δ𝐬t\)\.\\tilde\{\\mathbf\{z\}\}\_\{t\}=g\_\{\\phi\}\(\\mathbf\{h\}\_\{t\},\\Delta\\mathbf\{s\}\_\{t\}\)\.The posterior code is available only during training\. The hard prior code is also evaluated during training through prior alignment and next\-state reconstruction, and it is the only code used at rollout\. We quantize the posterior vector with the learned codebook𝒞=\{𝐜j\}j=1K\\mathcal\{C\}=\\\{\\mathbf\{c\}\_\{j\}\\\}\_\{j=1\}^\{K\}, whereK=128K=128and𝐜j∈ℝ64\\mathbf\{c\}\_\{j\}\\in\\mathbb\{R\}^\{64\}: ktpost=argminj∈\{1,…,K\}‖𝐳~t−𝐜j‖22,𝐳tpost=𝐜ktpost\.k\_\{t\}^\{\\mathrm\{post\}\}=\\arg\\min\_\{j\\in\\\{1,\\ldots,K\\\}\}\\\|\\tilde\{\\mathbf\{z\}\}\_\{t\}\-\\mathbf\{c\}\_\{j\}\\\|\_\{2\}^\{2\},\\qquad\\mathbf\{z\}\_\{t\}^\{\\mathrm\{post\}\}=\\mathbf\{c\}\_\{k\_\{t\}^\{\\mathrm\{post\}\}\}\.The VQ objective is ℒVQ=‖sg\[𝐳~t\]−𝐳tpost‖22\+β‖𝐳~t−sg\[𝐳tpost\]‖22,\\mathcal\{L\}\_\{\\mathrm\{VQ\}\}=\\\|\\mathrm\{sg\}\[\\tilde\{\\mathbf\{z\}\}\_\{t\}\]\-\\mathbf\{z\}\_\{t\}^\{\\mathrm\{post\}\}\\\|\_\{2\}^\{2\}\+\\beta\\\|\\tilde\{\\mathbf\{z\}\}\_\{t\}\-\\mathrm\{sg\}\[\\mathbf\{z\}\_\{t\}^\{\\mathrm\{post\}\}\]\\\|\_\{2\}^\{2\},wheresg\[⋅\]\\mathrm\{sg\}\[\\cdot\]is the stop\-gradient operator\. The assignment induces the one\-hot posteriorqϕ\(kt∣⋅\)=𝟏\{kt=ktpost\}q\_\{\\phi\}\(k\_\{t\}\\mid\\cdot\)=\\mathbf\{1\}\\\{k\_\{t\}=k\_\{t\}^\{\\mathrm\{post\}\}\\\}\. The prior predicts this assignment from𝐡t\\mathbf\{h\}\_\{t\}alone: ℒprior=DKL\(qϕ\(kt∣⋅\)∥pθ\(kt∣𝐡t\)\)=−logpθ\(ktpost∣𝐡t\)\.\\mathcal\{L\}\_\{\\mathrm\{prior\}\}=D\_\{\\mathrm\{KL\}\}\\\!\\left\(q\_\{\\phi\}\(k\_\{t\}\\mid\\cdot\)\\,\\\|\\,p\_\{\\theta\}\(k\_\{t\}\\mid\\mathbf\{h\}\_\{t\}\)\\right\)=\-\\log p\_\{\\theta\}\(k\_\{t\}^\{\\mathrm\{post\}\}\\mid\\mathbf\{h\}\_\{t\}\)\.The prior\-alignment arrow in Fig\.[1](https://arxiv.org/html/2611.24999#acmlabel1)denotes this cross\-entropy\. The posterior discovers transition\-informed codes from observed futures, while the prior predicts them using only pre\-transition information\. ### 3\.5\.Transition\-Conditioned Macro Behavior Prediction The world model predicts both the next state and the next macro behavior\. The latent code summarizes how the current profile, environment–behavior state, and macro behavior are associated with the next transition\. We compute the transition\-conditioned logits as ℓt\+1trans=fθtrans\(𝐡t,𝐳t,𝐬^t\+1\)\.\\ell\_\{t\+1\}^\{\\mathrm\{trans\}\}=f\_\{\\theta\}^\{\\mathrm\{trans\}\}\(\\mathbf\{h\}\_\{t\},\\mathbf\{z\}\_\{t\},\\hat\{\\mathbf\{s\}\}\_\{t\+1\}\)\.A context branch provides a complementary direct prediction: ℓt\+1ctx=fθctx\-macro\(𝐡t,𝐬t\)\.\\ell\_\{t\+1\}^\{\\mathrm\{ctx\}\}=f\_\{\\theta\}^\{\\mathrm\{ctx\\text\{\-\}macro\}\}\(\\mathbf\{h\}\_\{t\},\\mathbf\{s\}\_\{t\}\)\.We combine the branches with a learned gate: γt=σ\(fθgate\(𝐡t,𝐳t,𝐬^t\+1\)\),ℓt\+1=γtℓt\+1trans\+\(1−γt\)ℓt\+1ctx\.\\gamma\_\{t\}=\\sigma\\\!\\left\(f\_\{\\theta\}^\{\\mathrm\{gate\}\}\(\\mathbf\{h\}\_\{t\},\\mathbf\{z\}\_\{t\},\\hat\{\\mathbf\{s\}\}\_\{t\+1\}\)\\right\),\\qquad\\ell\_\{t\+1\}=\\gamma\_\{t\}\\ell\_\{t\+1\}^\{\\mathrm\{trans\}\}\+\(1\-\\gamma\_\{t\}\)\\ell\_\{t\+1\}^\{\\mathrm\{ctx\}\}\.This design makes macro prediction transition\-aware without requiring the latent branch when the current context is already sufficient\. ### 3\.6\.Joint Training Objective We jointly train state prediction, vector quantization, prior alignment, and macro behavior prediction: ℒ=λsℒstate\+λVQℒVQ\+λzℒprior\+λaℒmacro\.\\mathcal\{L\}=\\lambda\_\{s\}\\mathcal\{L\}\_\{\\mathrm\{state\}\}\+\\lambda\_\{\\mathrm\{VQ\}\}\\mathcal\{L\}\_\{\\mathrm\{VQ\}\}\+\\lambda\_\{z\}\\mathcal\{L\}\_\{\\mathrm\{prior\}\}\+\\lambda\_\{a\}\\mathcal\{L\}\_\{\\mathrm\{macro\}\}\.During training, the shared decoder forms𝐬^t\+1r=fθwm\(𝐡t,𝐳tr\)\\hat\{\\mathbf\{s\}\}\_\{t\+1\}^\{r\}=f\_\{\\theta\}^\{\\mathrm\{wm\}\}\(\\mathbf\{h\}\_\{t\},\\mathbf\{z\}\_\{t\}^\{r\}\)forr∈\{post,prior\}r\\in\\\{\\mathrm\{post\},\\mathrm\{prior\}\\\}\. The equally weighted objective is ℒstate=∑r∈\{post,prior\}‖𝐬^t\+1r−𝐬t\+1‖22\.\\mathcal\{L\}\_\{\\mathrm\{state\}\}=\\sum\_\{r\\in\\\{\\mathrm\{post\},\\mathrm\{prior\}\\\}\}\\\|\\hat\{\\mathbf\{s\}\}\_\{t\+1\}^\{r\}\-\\mathbf\{s\}\_\{t\+1\}\\\|\_\{2\}^\{2\}\.The alignment loss trains the prior distribution despite hard selection, while prior reconstruction exposes the decoder to the rollout code\. At rollout, only the prior\-code prediction is used\. For the observed next macro behaviorat\+1⋆a\_\{t\+1\}^\{\\star\}, ℒmacro=−logpθ\(at\+1⋆∣𝐪i,𝐬t,at\)\.\\mathcal\{L\}\_\{\\mathrm\{macro\}\}=\-\\log p\_\{\\theta\}\(a\_\{t\+1\}^\{\\star\}\\mid\\mathbf\{q\}\_\{i\},\\mathbf\{s\}\_\{t\},a\_\{t\}\)\.The gated decoder is trained throughℒmacro\\mathcal\{L\}\_\{\\mathrm\{macro\}\}applied to the combined logits\. All variants use the same class\-balanced macro loss\. The VQ\-based variants additionally use a lightweight code\-usage regularizer to mitigate codebook collapse\. ### 3\.7\.Simulation\-Grounded Execution After a macro behavior is predicted, the execution policy realizes it as a short\-horizon trajectory segment\. The macro behavior specifies the high\-level intent, while the micro controller instantiates that intent under the current simulator state\. Letξt:t\+H\\xi\_\{t:t\+H\}denote the executable segment over horizonHH\. The execution policy is πωexec\(ξt:t\+H∣𝐪i,𝐬t,at\)\.\\pi\_\{\\omega\}^\{\\mathrm\{exec\}\}\(\\xi\_\{t:t\+H\}\\mid\\mathbf\{q\}\_\{i\},\\mathbf\{s\}\_\{t\},a\_\{t\}\)\.Here,𝐪i\\mathbf\{q\}\_\{i\}preserves participant\-level conditioning,𝐬t\\mathbf\{s\}\_\{t\}grounds execution in the simulator, andata\_\{t\}specifies the macro intent\. We train the execution policy with DPO\([Rafailov et al\., 2023](https://arxiv.org/html/2611.24999#bib.bib13)\)\. Under the same context, pairs are ranked by deviation from the corresponding human trajectory:ξ\+\\xi^\{\+\}is the closer match andξ−\\xi^\{\-\}has substantially larger deviation\. The objective is \(1\)ℒDPO=−logσ\(β\[\\displaystyle\\mathcal\{L\}\_\{\\mathrm\{DPO\}\}=\-\\log\\sigma\\Bigg\(\\beta\\Bigg\[logπωexec\(ξ\+∣𝐪i,𝐬t,at\)πref\(ξ\+∣𝐪i,𝐬t,at\)\\displaystyle\\log\\frac\{\\pi\_\{\\omega\}^\{\\mathrm\{exec\}\}\(\\xi^\{\+\}\\mid\\mathbf\{q\}\_\{i\},\\mathbf\{s\}\_\{t\},a\_\{t\}\)\}\{\\pi\_\{\\mathrm\{ref\}\}\(\\xi^\{\+\}\\mid\\mathbf\{q\}\_\{i\},\\mathbf\{s\}\_\{t\},a\_\{t\}\)\}−\\displaystyle\-logπωexec\(ξ−∣𝐪i,𝐬t,at\)πref\(ξ−∣𝐪i,𝐬t,at\)\]\),\\displaystyle\\log\\frac\{\\pi\_\{\\omega\}^\{\\mathrm\{exec\}\}\(\\xi^\{\-\}\\mid\\mathbf\{q\}\_\{i\},\\mathbf\{s\}\_\{t\},a\_\{t\}\)\}\{\\pi\_\{\\mathrm\{ref\}\}\(\\xi^\{\-\}\\mid\\mathbf\{q\}\_\{i\},\\mathbf\{s\}\_\{t\},a\_\{t\}\)\}\\Bigg\]\\Bigg\),whereπref\\pi\_\{\\mathrm\{ref\}\}is the reference execution policy andβ\\betacontrols preference strength\. ## 4\.Experiments Experiments examine whether transition\-informed supervision improves behavior prediction for unseen participants, how continuous next\-state modeling, discretization, and prior alignment contribute, and whether the same profile–state–behavior factorization can be operationalized in a structurally different simulator\. VEET provides the primary human\-participant evaluation; AgentSociety provides a synthetic cross\-simulator test\. ### 4\.1\.Experimental Setup #### Simulation Environment\. We evaluate our framework on two complementary simulation platforms:AgentSociety\([Piao et al\., 2025](https://arxiv.org/html/2611.24999#bib.bib2)\)andVEET, a self\-built Unity\-based behavioral simulator\. AgentSociety is a synthetic, LLM\-driven city\-scale multi\-agent simulation platform in which agents maintain internal memory, update needs and internal simulator states, generate daily activity plans, and execute behaviors through modular action blocks\. It provides a synthetic testbed for operationalizing profile\-conditioned behavior under a different state and action interface\. VEET is a Unity\-based simulator designed to study behavioral trade\-offs between reward acquisition and anxiety\- or effort\-related pressure\. Further environment details appear in Section[B](https://arxiv.org/html/2611.24999#A2)\. Participants make sequential navigation and task decisions in a controlled interactive environment\. #### AgentSociety Simulation\. We use AgentSociety as a synthetic testbed and adapt it to our hierarchical formulation by injecting questionnaire\-derived psychological profiles into each LLM\-controlled agent and constraining its behavior to five macro actions:work,consume,rest,leisure, andmove\. Each agent is simulated over a full daily schedule rather than through isolated one\-step queries\. At each decision point, the LLM observes the agent’s current needs, profile\-derived traits, location, and recent context, and selects a macro action that is executed over the corresponding simulation interval\. This yields temporally grounded macro\-action traces that can be directly compared with our hierarchical policy models\. Implementation details of this adaptation appear in Section[A](https://arxiv.org/html/2611.24999#A1)\. \(a\)VEET human–agent trajectory comparison \(b\)AgentSociety observed vs\. predicted macro distributions Figure 2\.Cross\-environment qualitative diagnostics\. Panel \(a\) compares human and agent trajectories for two held\-out VEET participants\. Panel \(b\) compares observed and gated VQ\-WM macro\-action distributions within AgentSociety\-specific questionnaire\-profile clusters; its compact labels C0–C3 correspond to A\-C0–A\-C3 in Table[4](https://arxiv.org/html/2611.24999#S4.T4)\. Here,nndenotes the number of held\-out macro\-action events rather than unique participants\. The panels visualize environment\-specific outcomes and are not pooled\.Panel \(a\) compares VEET human and agent trajectories\. Panel \(b\) compares observed and predicted AgentSociety macro\-action distributions across environment\-specific questionnaire\-profile clusters\.Figure 3\.Exploratory associations between participant\-level prior\-code usage frequencies and seven psychometric scales\. Cells report Spearmanρ\\rho;∗\*denotes BH–FDRq<0\.05q<0\.05across 133 tests and†\\daggerdenotes uncorrectedp<0\.05p<0\.05\. No association survives FDR correction\. The median pairwise sample size is 297, and codes are ordered by maximum absolute correlation\.A diverging heatmap of Spearman correlations between 19 prior VQ\-code usage frequencies and seven questionnaire scales\. Correlations are small, and none survives false\-discovery\-rate correction\. #### Dataset and Subjects\. We use questionnaire profiles from 313 VEET participants as the common subject representation across both environments\. In VEET, these participants provide questionnaire profiles and behavioral trajectories for training the state\-transition world model and macro\-behavior predictor\. Post\-level subjective ratings are used only as auxiliary targets in the Direct SFT/BC baseline and are not modeled as dynamic affective states by the proposed world model\. For AgentSociety, the same profile vectors initialize synthetic LLM\-controlled agents in the modified simulator\. Agents are rolled out over a one\-day schedule starting at 08:00 with a 5\-minute base simulation resolution, and trajectories are logged event\-wise when actions are committed\. Each LLM decision is constrained to the structured macro vocabulary\{work,consume,rest,leisure,move\}\\\{\\texttt\{work\},\\texttt\{consume\},\\texttt\{rest\},\\texttt\{leisure\},\\texttt\{move\}\\\}, yielding profile\-conditioned daily behavior traces\. #### Subject\-Level OOD Evaluation\. We treat the human participant as the domain factor and use a strictly subject\-disjoint split of240 training,25 validation, and48 test participants\. Each participant, questionnaire profile, and associated trajectory belongs to exactly one split; there is no participant overlap across training, validation, and test\. This protocol therefore tests whether the behavioral predictor generalizes to unseen people, rather than measuring performance on new trajectories from previously observed people\. Table[1](https://arxiv.org/html/2611.24999#S4.T1)reports this subject\-level OOD evaluation\. #### Implementation Details\. The trajectory encoderEϕE\_\{\\phi\}is a two\-layer transformer with hidden dimension 128 and four attention heads\. The VQ codebook containsK=128K=128entries of dimension 64, updated via exponential moving average with decay 0\.99\. The profile encoderEψqE\_\{\\psi\}^\{q\}is a two\-layer MLP with hidden dimension 64 and ReLU activations\. The macro decoder uses a gated two\-branch architecture: one branch conditions on the transition code and predicted next state, and the other on current context; the logits from the two branches are combined with a learned scalar gate\. All models are trained with Adam\(lr=3×10−4,β1=0\.9,β2=0\.999\)\(\\mathrm\{lr\}=3\\times 10^\{\-4\},\\beta\_\{1\}=0\.9,\\beta\_\{2\}=0\.999\)for 100 epochs with batch size 256\. The objective weights are set toλs=1\.0\\lambda\_\{s\}=1\.0,λVQ=0\.25\\lambda\_\{\\mathrm\{VQ\}\}=0\.25,λz=0\.1\\lambda\_\{z\}=0\.1, andλa=1\.0\\lambda\_\{a\}=1\.0\. All variants use the same class\-balanced macro loss\. The VQ\-based variants additionally use the code\-usage regularizer described in Section[3](https://arxiv.org/html/2611.24999#S3); the gated branches are optimized through the combined macro loss\. All experiments are run on a single NVIDIA A100 GPU\. #### Evaluation Metrics\. We report Top\-1 and Top\-3 accuracy for next\-macro\-behavior prediction\. LetPdP\_\{d\}andPpP\_\{p\}denote the empirical and predicted macro\-action distributions\. Macro KL isDKL\(Pd∥Pp\)D\_\{\\mathrm\{KL\}\}\(P\_\{d\}\\\|P\_\{p\}\), and macro\-distribution similarity is1−JSD\(Pd,Pp\)/log21\-\\mathrm\{JSD\}\(P\_\{d\},P\_\{p\}\)/\\log 2\. These metrics evaluate distributional alignment beyond the dominant class; the JSD\-based measure does not by itself establish long\-range temporal dependence\. We also report the number of unique VQ codes activated by the prior as a diagnostic of codebook collapse\. Table 1\.Main results under subject\-level OOD evaluation on VEET and AgentSociety\. The 240/25/48 train/validation/test participants are mutually exclusive, so the test set evaluates generalization to unseen people rather than new trajectories from people observed during training\. Macro Top\-1/Top\-3 evaluate next macro\-action prediction; Macro KL isDKL\(Pdata\(a\)∥Ppred\(a\)\)D\_\{\\mathrm\{KL\}\}\(P\_\{\\mathrm\{data\}\}\(a\)\\\|P\_\{\\mathrm\{pred\}\}\(a\)\), and Dist\. Sim\. is the macro\-action distribution similarity1−JSD\(Pdata,Ppred\)/log21\-\\mathrm\{JSD\}\(P\_\{\\mathrm\{data\}\},P\_\{\\mathrm\{pred\}\}\)/\\log 2\. For Ours on VEET, the second line reports 95% participant\-level bootstrap confidence intervals over all 48 test participants \(1,000 resamples; seed 2026\)\. ### 4\.2\.Unseen\-Participant Behavioral Prediction Table[1](https://arxiv.org/html/2611.24999#S4.T1)evaluates the complete 48\-participant test set, with no participant, profile, or trajectory shared with training\. On VEET, the profile\-conditioned world model reaches 0\.904 Top\-1 and 0\.989 Top\-3 accuracy, compared with 0\.754 and 0\.890 for direct SFT/BC\. It also reduces Macro KL to 0\.0033 and reaches 0\.999 macro\-distribution similarity\. The participant\-level 95% bootstrap interval for Top\-1 is \[0\.895, 0\.913\], with the remaining intervals reported directly in the table\. Together, these results show that the learned transition representation supports macro\-behavior prediction for previously unseen participants; they do not establish causal psychological dynamics\. Table 2\.VEET latent representation ablation\. ### 4\.3\.Contribution of Transition Supervision Table[2](https://arxiv.org/html/2611.24999#S4.T2)separates continuous next\-state prediction, discrete transition modeling, and rollout\-time prior alignment\. The largest pointwise difference is between direct SFT/BC in Table[1](https://arxiv.org/html/2611.24999#S4.T1)\(0\.754 Top\-1\) and the continuous world model \(0\.901\), associating explicit next\-state supervision with most of the Top\-1 gain\. Because these variants also differ architecturally, this comparison is evidence of predictive utility rather than a controlled causal attribution\. Vector quantization changes Top\-1 only slightly \(0\.905\) but improves Top\-3 from 0\.952 to 0\.981 and reduces Macro KL from 0\.012 to 0\.0078\. Prior alignment further lowers Macro KL to 0\.0033 and raises Top\-3 to 0\.989 while preserving Top\-1\. Thus, continuous transition supervision supplies the principal pointwise gain, whereas the discrete prior–posterior design mainly improves behavioral coverage and distributional alignment while making the transition\-informed representation available before the future is observed\. A profile\-cluster diagnostic tests whether this aggregate result is concentrated in a small questionnaire subgroup\. VEET Top\-1 remains between 0\.891 and 0\.913 across the reported groups, although teleport recall is lower and more variable \(0\.13–0\.26\)\. Because cluster sizes are limited, we treat these differences as descriptive boundaries rather than subgroup or causal effects \(Table[3](https://arxiv.org/html/2611.24999#S4.T3)\)\. Table 3\.Descriptive held\-out VEET performance by environment\-specific questionnaire\-profile cluster\. ### 4\.4\.What the Latent Modes Represent The latent variable is intended to summarize transition\-level predictive structure, not to encode a static personality category\. We therefore examine 19 analyzed prior codes against seven psychometric scales \(133 tests\)\. No code–scale association survives BH–FDR correction, and nominal effects are small \(\|ρ\|≈0\.11\|\\rho\|\\approx 0\.11–0\.190\.19\)\. This result does not prove disentanglement, but it is consistent with the transition modes not being reducible to individual questionnaire dimensions \(Figure[3](https://arxiv.org/html/2611.24999#acmlabel3)\)\. At the uncorrected level, several small associations provide exploratory structure without supporting psychometric labels\. C90 is positively associated with PSS, STAI\-T, BDI, and STAI\-S; C23 and C91 show related stress–anxiety patterns; C73 is positively associated with AMI; and C88, C74, and C48 are negatively associated with STAI\-S\. Code\-use entropy and the number of unique codes likewise yield no FDR\-significant scale association\. We therefore treat these patterns as hypothesis\-generating diagnostics of transition\-mode usage, not as evidence that individual codes correspond to traits, symptoms, or clinical categories\. ### 4\.5\.Synthetic Cross\-Simulator Operationalization AgentSociety tests whether the same profile–state–behavior factorization can be operationalized in a synthetic LLM\-based simulator with a different state space and action interface\. Ours obtains the highest reported Top\-1 accuracy \(0\.595\) and maintains strong macro\-distribution alignment in Table[1](https://arxiv.org/html/2611.24999#S4.T1)\. Figure[2\(b\)](https://arxiv.org/html/2611.24999#S4.F2.sf2)further compares observed and predicted cluster\-level action mixtures\. The rollouts preserve differences among work, consumption, leisure, rest, and movement across profile groups, while the prediction tracks the dominant cluster\-level patterns\. We interpret this as a synthetic cross\-simulator operationalization, not as a second independent validation of human psychological mechanisms; the AgentSociety agents are initialized from questionnaire\-derived profiles\. For descriptive diagnostics, we fit and label the questionnaire\-profile clusters separately within each environment; their identifiers are therefore environment\-specific and are not comparable across Tables[3](https://arxiv.org/html/2611.24999#S4.T3)and[4](https://arxiv.org/html/2611.24999#S4.T4)\. AgentSociety implementation details appear in Appendix[A](https://arxiv.org/html/2611.24999#A1)\. Table 4\.Descriptive held\-out AgentSociety performance by environment\-specific questionnaire\-profile cluster\.Table[4](https://arxiv.org/html/2611.24999#S4.T4)further tests whether cross\-simulator performance is concentrated in one questionnaire\-defined subgroup\. Top\-1 varies from 0\.558 to 0\.613 across the four clusters, while leisure recall ranges from 0\.16 to 0\.27\. The mid\-profile cluster is comparatively harder, whereas the remaining clusters have similar aggregate accuracy despite different stress, BDI, and hedonic profiles\. Because these groups contain 10–14 held\-out participants, we interpret the differences as descriptive diagnostics rather than causal subgroup effects\. The narrow performance range indicates that the aggregate AgentSociety result is not driven by a single profile cluster\. Table 5\.Micro\-control results for SFT/BC and macro\-conditioned DPO\.Table[5](https://arxiv.org/html/2611.24999#S4.T5)defines Completion as the fraction of rollouts completing the requested macro action, Coin gain as the mean number of VEET coin rewards per rollout, First\-act\. KL as the divergence between model and human first\-action distributions, and Move\. gap as the absolute difference in overallmovefrequency\. DPO improves completion and reduces both gaps; on VEET, it also increases coin gain\. ## 5\.Limitations and Ethical Considerations The VEET data were collected by the original study team under institutional human\-subjects oversight at Emory University, with informed consent from all participants\. This work uses only de\-identified data and involved no new participant recruitment or contact; access was restricted to authorized researchers\. Our evaluation establishes the downstream predictive utility of transition supervision; it does not independently assess calibrated multi\-step state\-transition fidelity\. The model does not contain a separately observed dynamic affect channel; its transition representation is learned from environmental and behavioral variables recorded by VEET\. The learned codes are behavior\-predictive summaries rather than causal, clinical, or diagnostic states, and the model is not intended for consequential decisions about individuals\. Participant\-disjoint evaluation addresses transfer to unseen people within this study design; validation across additional populations and instruments remains future work\. Responsible use requires de\-identification, restricted access, and human oversight\. ## 6\.Conclusion We study transition\-informed supervision for behavior prediction in unseen participants\. The world model separates stable profiles, current states, and behavior: a posterior extracts codes from observed changes, while a prior predicts them before the future is available\. VEET results associate next\-state modeling with pointwise accuracy and discretization and prior alignment with behavioral coverage and distributional alignment\. AgentSociety provides a synthetic cross\-simulator operationalization\. These results support the world model as a predictive inductive bias rather than a validated psychological mechanism\. ## 7\.Generative AI Usage Generative AI tools assisted manuscript organization and language editing, LaTeX and figure preparation, and development and debugging of analysis code\. ## References - Abdulhaiet al\.\(2025\)M\. Abdulhai, R\. Cheng, D\. Clay, T\. Althoff, S\. Levine, and N\. JaquesConsistently simulating human personas with multi\-turn reinforcement learning\.External Links:2511\.00222,[Link](https://arxiv.org/abs/2511.00222)Cited by:[§1](https://arxiv.org/html/2611.24999#S1.p2.1),[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px2.p1.1)\. - Brookeset al\.\(2019\)J\. Brookes, M\. Warburton, M\. Alghadier, M\. Mon\-Williams, and F\. MushtaqStudying human behavior with virtual reality: the unity experiment framework\.Behavior Research Methods52\(2\),pp\. 455–463\.External Links:ISSN 1554\-3528,[Link](http://dx.doi.org/10.3758/s13428-019-01242-0),[Document](https://dx.doi.org/10.3758/s13428-019-01242-0)Cited by:[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px1.p1.1)\. - Duanet al\.\(2025\)Y\. Duan, Y\. Tang, X\. Bai, K\. Chen, J\. Li, and M\. ZhangThe power of personality: a human simulation perspective to investigate large language model agents\.External Links:2502\.20859,[Link](https://arxiv.org/abs/2502.20859)Cited by:[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px2.p1.1)\. - Hafneret al\.\(2020\)D\. Hafner, T\. Lillicrap, J\. Ba, and M\. NorouziDream to control: learning behaviors by latent imagination\.External Links:1912\.01603,[Link](https://arxiv.org/abs/1912.01603)Cited by:[§1](https://arxiv.org/html/2611.24999#S1.p4.1),[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px3.p1.1)\. - Hafneret al\.\(2022\)D\. Hafner, T\. Lillicrap, M\. Norouzi, and J\. BaMastering atari with discrete world models\.External Links:2010\.02193,[Link](https://arxiv.org/abs/2010.02193)Cited by:[§1](https://arxiv.org/html/2611.24999#S1.p4.1),[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px3.p1.1)\. - Hafneret al\.\(2023\)D\. Hafner, J\. Pasukonis, J\. Ba, and T\. LillicrapMastering diverse domains through world models\.arXiv preprint arXiv:2301\.04104\.Cited by:[§1](https://arxiv.org/html/2611.24999#S1.p4.1),[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px3.p1.1)\. - Hewittet al\.\(2025\)S\. R\. C\. Hewitt, A\. Norbury, Q\. J\. M\. Huys, and T\. U\. HauserDay\-to\-day fluctuations in motivation drive effort\-based decision\-making\.Proceedings of the National Academy of Sciences122\(12\),pp\. e2417964122\.External Links:[Document](https://dx.doi.org/10.1073/pnas.2417964122),[Link](https://www.pnas.org/doi/abs/10.1073/pnas.2417964122),https://www\.pnas\.org/doi/pdf/10\.1073/pnas\.2417964122Cited by:[§1](https://arxiv.org/html/2611.24999#S1.p1.1),[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px1.p1.1)\. - Kozlov and Johansen \(2010\)M\. D\. Kozlov and M\. K\. JohansenReal behavior in virtual environments: psychology experiments in a simple virtual\-reality paradigm using video games\.Cyberpsychology, Behavior, and Social Networking13,pp\. 711 – 714\.External Links:[Link](https://api.semanticscholar.org/CorpusID:9400778)Cited by:[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px1.p1.1)\. - Parket al\.\(2026\)J\. S\. Park, C\. Q\. Zou, J\. Kamphorst, N\. Egan, A\. Shaw, B\. M\. Hill, C\. Cai, M\. R\. Morris, P\. Liang, R\. Willer, and M\. S\. BernsteinLLM agents grounded in self\-reports enable general\-purpose simulation of individuals\.External Links:2411\.10109,[Link](https://arxiv.org/abs/2411.10109)Cited by:[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px1.p1.1)\. - Piaoet al\.\(2025\)J\. Piao, Y\. Yan, J\. Zhang, N\. Li, J\. Yan, X\. Lan, Z\. Lu, Z\. Zheng, J\. Y\. Wang, D\. Zhou,et al\.Agentsociety: large\-scale simulation of llm\-driven generative agents advances understanding of human behaviors and society\.Cited by:[§1](https://arxiv.org/html/2611.24999#S1.p2.1),[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px2.p1.1),[§4\.1](https://arxiv.org/html/2611.24999#S4.SS1.SSS0.Px1.p1.1)\. - Rafailovet al\.\(2023\)R\. Rafailov, A\. Sharma, E\. Mitchell, C\. D\. Manning, S\. Ermon, and C\. FinnDirect preference optimization: your language model is secretly a reward model\.Advances in neural information processing systems36,pp\. 53728–53741\.Cited by:[§3\.7](https://arxiv.org/html/2611.24999#S3.SS7.p2.1)\. - Ren and Xu \(2025\)M\. Ren and W\. XuThe impact of big five personality traits on ai agent decision\-making in public spaces: a social simulation study\.External Links:2503\.15497,[Link](https://arxiv.org/abs/2503.15497)Cited by:[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px2.p1.1)\. - Wanget al\.\(2025\)Y\. Wang, Y\. Chen, F\. Zhong, L\. Ma, and Y\. WangSimulating human\-like daily activities with desire\-driven autonomy\.External Links:2412\.06435,[Link](https://arxiv.org/abs/2412.06435)Cited by:[§1](https://arxiv.org/html/2611.24999#S1.p2.1),[§2](https://arxiv.org/html/2611.24999#S2.SS0.SSS0.Px2.p1.1)\. ## Appendix AAgentSociety ### A\.1\.Modifications We use AgentSociety as a synthetic testbed and modify it to expose a profile\-conditioned hierarchical action interface\. The modification consists of four components: profile injection, need\-dynamics modification, block\-level action restriction, and trajectory logging\. Profile injection\.For each subject, we first convert the questionnaire vector into anAgentPsychProfile\. The profile contains initial need satisfactions, emotion values, personality text, and three composite behavioral dimensions: hedonic capacity, effort capacity, and stress level\. After the simulator creates the nativeSocietyAgentinstances, we patch each live agent by writing these values into its memory state\. Specifically, we initializehunger\_satisfaction,energy\_satisfaction,safety\_satisfaction,social\_satisfaction, and the addedpleasure\_satisfactionfield, together with the agent’s emotion state, emotion types, background text, and initial thought\. Need\-dynamics modification\.We extend the original need state with a pleasure\-related dimension\. This dimension is initialized from questionnaire\-derived hedonic traits, especially measures related to anhedonia and momentary pleasure\. We also patch theNeedsBlockparameters after agent construction\. Questionnaire traits scale the decay rates of hunger, energy, safety, social, and pleasure needs through multiplicative factorsαH,αD,αP,αC,αL\\alpha\_\{H\},\\alpha\_\{D\},\\alpha\_\{P\},\\alpha\_\{C\},\\alpha\_\{L\}\. Anxiety\- and hedonic\-related traits further shift the safety and pleasure thresholds\. This makes the same environment state induce different dominant needs for different profile types\. For example, high stress can make safety needs trigger earlier, while low hedonic capacity can make pleasure satisfaction decay faster and trigger leisure\-seeking behavior sooner\. Social\-behavior suppression\.Because our experiments focus on individual profile\-conditioned behavior rather than interpersonal dynamics, we suppress the original social pathway\. We remove social behavior from the plan guidance options, replace the detailed planning prompt with a no\-social version, and set the social satisfaction state to a high value while making its decay nearly zero\. We also remove any registeredSocialBlockfrom the agent dispatcher when present\. This prevents the agent from spending trajectory steps on messaging, calling, or meeting other agents, and ensures that logged behaviors correspond to individual daily activities\. Structured macro\-action interface\.The original AgentSociety planner can generate free\-form natural\-language actions\. We replace this with a five\-category macro\-action vocabulary:𝒜=\{work,consume,rest,leisure,move\}\\mathcal\{A\}=\\\{\\texttt\{work\},\\texttt\{consume\},\\texttt\{rest\},\\texttt\{leisure\},\\texttt\{move\}\\\}\. Internally, these correspond to coarse plan typeswork,consume,sleep,leisure, andmobility\. The detailed plan prompt is constrained so that each plan step must use exactly one of these coarse types and must avoid fine\-grained sub\-actions\. We additionally introduce or expose dedicated rest and leisure execution paths so thatrestandleisurecorrespond to executed behavior modules rather than merely textual descriptions\. The macro\-action prior is profile\-conditioned\. The current dominant need provides a base preference over macro actions: hunger favors consumption, tiredness favors rest, safety favors work, and pleasure favors leisure\. The composite profile dimensions then shift these logits: hedonic capacity increases the tendency toward consumption and leisure, effort capacity increases work and mobility, and stress increases rest while suppressing consumption, leisure, and movement\. The resulting logits define a profile\-conditioned macro distribution used both for simulation analysis and as the behavioral abstraction aligned with our hierarchical policy models\. Algorithm 1AgentSociety profile\-conditioned simulation1:Questionnaire profiles \{qi\}\\\{q\_\{i\}\\\}, selected subjects 𝒮\\mathcal\{S\} 2:for i∈𝒮i\\in\\mathcal\{S\}do 3: pi←QuestionnaireToProfile\(qi\)p\_\{i\}\\leftarrow\\textsc\{QuestionnaireToProfile\}\(q\_\{i\}\) 4:Create AgentSociety agent aia\_\{i\} 5:Inject pip\_\{i\}into memory needs, emotions, personality, and thought 6:PatchNeedsBlockdecay rates and thresholds 7:Remove social guidance and disableSocialBlock 8:Restrict plan steps to \{work,consume,rest,leisure,move\}\\\{\\texttt\{work\},\\texttt\{consume\},\\texttt\{rest\},\\texttt\{leisure\},\\texttt\{move\}\\\} 9:endfor 10:foreach simulation time step ttdo 11:foreach agent aia\_\{i\}do 12:Observe current simulator state, location, memory, and profile 13:Compute profile\-conditioned macro prior 14:LLM selects a structured macro action and generates a plan 15:Execute the corresponding AgentSociety behavior block 16:Log state, profile, macro action, probabilities, and description 17:endfor 18:endfor Execution and logging\.Each LLM\-controlled agent is simulated over a daily schedule\. At each discrete simulation time point, the agent observes its current need state, emotion, thought, profile text, location, and recent memory context\. The LLM selects a guidance option and expands it into a short structured plan whose steps belong to the allowed macro categories\. During post\-processing, each executed memory stream event is mapped back to one of the five macro actions\. We log the simulation tick, current need state, satisfaction values, questionnaire profile, macro action, macro\-action logits and probabilities, and the corresponding micro\-action description\. This produces event\-level trajectories that are temporally grounded in the AgentSociety simulator and directly comparable to the macro\-action sequences used by our hierarchical world\-model and policy components\. #### AgentSociety trajectory generation details\. For AgentSociety, each rollout uses the VEET questionnaire\-derived profile vector𝐪i\\mathbf\{q\}\_\{i\}to initialize the agent’s native need and emotion fields\. The profile determines hedonic capacity, effort capacity, perceived stress, and need\-specific decay or threshold parameters\. Simulations begin at 08:00 and run over a one\-day schedule\. AgentSociety advances withticks\_per\_step=300, corresponding to a 5\-minute simulation resolution\. However, logged trajectories are event\-based rather than uniformly sampled: a record is written when the agent commits an action in its memory stream\. As a result, logged behavioral intervals are variable, typically ranging from 5 to 30 simulated minutes\. At each decision point, the LLM observes the agent’s current native simulator state, location, profile\-derived traits, and recent memory context\. It first selects a profile\-conditioned guidance option and then expands it into a structured plan of up to six coarse steps\. Each step is constrained to one of five macro\-action categories:\{work,consume,rest,leisure,move\}\\\{\\texttt\{work\},\\texttt\{consume\},\\texttt\{rest\},\\texttt\{leisure\},\\texttt\{move\}\\\}\. These event\-level macro\-action traces form the AgentSociety dataset used for profile\-conditioned behavior modeling\. ## Appendix BVEET Explanation ### B\.1\.VEET Environment and Training State Details #### VEET environment\. VEET is an embodied Unity\-based behavioral task in which participants navigate through a multi\-room virtual environment, collect coins, move between rooms through doors, and occasionally use teleportation\. The environment records both low\-level motor behavior and task\-level progress variables\. Each participant also provides questionnaire responses measuring affective and motivational traits\. These questionnaire responses define a subject profile vector𝐪i\\mathbf\{q\}\_\{i\}, which conditions our models throughout training\. #### Observation state\. At each time step, we convert the raw Unity observation into a flat numerical state vector𝐬t\\mathbf\{s\}\_\{t\}\. The state contains four groups of information: \(1\) self\-motion and position, including player position, rotation, movement distance, and wall distances; \(2\) room context, including current room, room effort, room reward, door positions, and distances to front and back doors; \(3\) object layout, including remaining coin counts and nearest\-coin relative positions/distances; and \(4\) task progress, including collected coins, points, button presses, teleport counts, elapsed time, room time, round time, and timing flags for post\-level subjective queries, but not the subjective rating values\. The raw state vector is normalized with the same feature scaler used by the behavioral cloning and reward\-model baselines\. When clean\-feature checkpoints are used, we apply the corresponding feature mask before normalization so that all models use an identical state interface\. #### Questionnaire conditioning\. Each subjectiiis associated with a questionnaire\-derived vector𝐪i\\mathbf\{q\}\_\{i\}\. Depending on the experiment,𝐪i\\mathbf\{q\}\_\{i\}is represented either as item\-level responses, compressed scale\-level features, or eight scale scores corresponding to PSS, SHAPS, BDI, AES, AMI, MAPSR, STAI\-S, and STAI\-T\. The same𝐪i\\mathbf\{q\}\_\{i\}is concatenated with the environment state or encoded as profile context in the policy and world\-model networks\. Thus, the model predicts behavior not only from the current physical state, but also from stable subject\-level psychometric traits\. #### Macro\-action labels\. We derive a six\-way macro\-action label from human replay trajectories\. These labels are inferred from geometric progress, movement speed, button presses, room effort, and short\-horizon events\. For example,go\_to\_coincorresponds to movement that reduces distance to nearby coins,go\_to\_next\_doorandgo\_to\_prev\_doorcorrespond to room\-transition\-directed behavior, andteleportcorresponds to explicit teleport usage\. This macro abstraction provides the high\-level action space for our hierarchical policy and world model\. #### Micro\-action space\. The lower\-level VEET controller uses the native discrete Unity action space\. Each action is represented as a triplet\(mt,rt,zt\)\(m\_\{t\},r\_\{t\},z\_\{t\}\), wheremtm\_\{t\}is movement intensity,rtr\_\{t\}is turn direction, andztz\_\{t\}indicates teleportation\. Non\-teleport actions enumerate movement and turn combinations, while a separate discrete action is reserved for teleport\. This gives a 13\-way micro\-action space\. The macro\-conditioned micro policy receives the normalized state𝐬t\\mathbf\{s\}\_\{t\}, the subject profile𝐪i\\mathbf\{q\}\_\{i\}, and an embedding of the current macro action, then outputs a distribution over these 13 low\-level actions\. #### Training transitions\. We distinguish the state\-world\-model and micro\-control interfaces\. The former uses 𝒟WM=\{\(𝐬t,𝐪i,atmacro,𝐬t\+1,at\+1macro\)\},\\mathcal\{D\}\_\{\\mathrm\{WM\}\}=\\\{\(\\mathbf\{s\}\_\{t\},\\mathbf\{q\}\_\{i\},a\_\{t\}^\{\\mathrm\{macro\}\},\\mathbf\{s\}\_\{t\+1\},a\_\{t\+1\}^\{\\mathrm\{macro\}\}\)\\\},whereas the micro controller uses 𝒟micro=\{\(𝐬t,𝐪i,atmacro,atmicro,𝐬t\+1\)\}\.\\mathcal\{D\}\_\{\\mathrm\{micro\}\}=\\\{\(\\mathbf\{s\}\_\{t\},\\mathbf\{q\}\_\{i\},a\_\{t\}^\{\\mathrm\{macro\}\},a\_\{t\}^\{\\mathrm\{micro\}\},\\mathbf\{s\}\_\{t\+1\}\)\\\}\.This separation prevents the observed micro action from entering the world\-model transition tuple\. For macro\-conditioned DPO, candidate trajectory segments are initialized from the same replay context and conditioned on the macro label\. We form preference pairs by comparing candidates with the corresponding human replay trajectory using progress signals such as coin gain, door progress, movement distance, room change, and teleport events\. The closer candidate is preferred, whereas a segment with substantially larger behavioral deviation is rejected\. #### Role in our framework\. VEET therefore provides both human demonstration trajectories and questionnaire\-conditioned subject profiles\. The trajectories supervise macro\-action prediction, micro\-action execution, and latent transition modeling, while the profiles allow the model to capture individual differences in strategy selection\. This makes VEET a controlled testbed for evaluating whether transition\-informed supervision supports profile\-conditioned behavior prediction beyond a majority\-action shortcut\.
Similar Articles
Mental World Modeling
The paper introduces Mental World Modeling (MWM), a framework that integrates hidden mental states as core components of world models, and presents MENTIS, a training-free baseline. Experiments with 8 LLM-based world models show explicit mental-state modeling is essential for predicting human decisions in situated scenarios.
World-in-Context: Interactive Video World Modeling with In-Context Conditioning
World-in-Context introduces an interactive video world modeling framework that unifies view, action, and camera controls as explicit tokens within a shared self-attention space via In-Context Conditioning, along with Interaction-Transport RoPE for temporal continuity. Its 5B model matches the navigation performance of the 22B EchoWM on WBench while achieving a 93.8% perspective-switching success rate versus 18.8% for existing 5B baselines.
WorldReward: Reward Modeling for Camera-Conditioned World Models
WorldReward introduces a vision-language reward model for camera-conditioned world models that unifies action-consistency and visual-quality evaluation through chunk decomposition and preference aggregation, outperforming existing methods like GPT-5.5 on benchmarks.
PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives
PlayWorld is a benchmark for evaluating interactive video world models using multi-modal agent players pursuing long-horizon objectives. It assesses geometry consistency, interaction fidelity, and state evolution, revealing that current models struggle with spatial consistency and persistent state evolution.
Training Object Permanence in World Models
This paper introduces WROP, a dataset for training object permanence in world models, and evaluates 14 video models, releasing PWM-WROP, a 16B world model that ranks among top continuation models.