betaFeedback

WorksComputer Science

PsyWorld: Profile-Conditioned World Modeling for Human Behavior Prediction

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.

ACM Reference Format:

Ziyuan Qin, Yang Qiao, Summer B. Frandsen, Michael T. Treadway, and Liang Zhao. 2027. PsyWorld: Profile-Conditioned World Modeling for Human Behavior Prediction. In Proceedings of KDD 2027 AI for Sciences Track (KDD ’27). ACM, New York, NY, USA, 11 pages.

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 [7]. 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 [10, 13], while consistency-oriented training reduces persona drift across interactions [1]. 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 [4, 5, 6]. 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.

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 [7]. Simulation platforms provide a controlled setting in which such dynamics can be observed and modeled: Unity-based experimental frameworks reproduce established behavioral effects [2, 8], while interview-conditioned generative agents can reproduce several empirical patterns in human responses [9]. 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 [10], whereas D2A maps personality profiles to desire-driven values for daily activity generation [13]. Other work has measured how Big Five traits alter LLM-agent decisions and performance [3, 12], or used consistency-based optimization to reduce persona drift across interactions [1]. 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 [4], while discrete and more general successors improved representation learning and cross-domain robustness [5, 6]. 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.

Method

Problem Formulation: Transition-Informed Behavioral Prediction

For participant ii, the stable questionnaire-derived profile is qi∈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, \mathbf{a}_t, \mathbf{s}_{t+1}, \mathbf{a}_{t+1}),

where st∈R49\mathbf{s}_t \in\mathbb{R}^{49} is the recorded environment–behavior state and at\mathbf{a}_t is the current macro behavior. Our primary empirical objective is

p(at+1∣qi,st,at)p(\mathbf{a}_{t+1} \mid\mathbf{q}_i, \mathbf{s}_t, \mathbf{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}, \mathbf{a}_{t+1} \mid\mathbf{q}_i, \mathbf{s}_t, \mathbf{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 to st+1s_{t+1}. This factorization treats the latent codes as predictive summaries of profile-conditioned state transitions, not as causal psychological or clinical constructs.

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.

Profile-Conditioned State World Model

We represent stable and time-varying information separately. The psychometric profile qiq_i is fixed for participant ii, whereas sts_t records the time-varying task, environmental, and behavioral state. The current macro behavior ata_t specifies the behavioral context under which the transition occurs. For AgentSociety,

AAS={work,consume,rest,leisure,move}.\mathcal{A}_{\mathrm{AS}} = \{\mathrm{work},\mathrm{consume},\mathrm{rest},\mathrm{leisure},\mathrm{move}\}.

The context encoder fuses these variables:

ht=fθctx(qi,st,at).h_t = f_{\theta}^{\mathrm{ctx}}(q_i, s_t, a_t).

Given a latent transition code ztz_t, the world-model decoder predicts the next environment–behavior state:

s^t+1=fθwm(ht,zt).\hat{s}_{t+1} = f_{\theta}^{\mathrm{wm}}(h_t, 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.

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 h_t) = \operatorname{Cat}(\operatorname{softmax}(W_z h_t)).

During training and rollout, we use the hard assignment ktprior=arg⁡max⁡kpθ(k∣ht)k_t^{\mathrm{prior}} = \arg\max_k p_{\theta}(k \mid h_t) and ztprior=cktpriorz_t^{\mathrm{prior}} = c_{k_t^{\mathrm{prior}}}. During training, the realized state change

Δst=st+1−st\Delta s_t = s_{t+1} - 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{z}_t = g_{\phi}(h_t, \Delta 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 codebook C={cj}j=1K\mathcal{C} = \{c_j\}_{j=1}^{K}, where K=128K = 128 and cj∈R64c_j \in\mathbb{R}^{64}:

ktpost=arg⁡min⁡j∈{1,…,K}∥z~t−cj∥22,ztpost=cktpost.k_t^{\mathrm{post}} = \arg\min_{j \in\{1,\ldots,K\}} \lVert\tilde{z}_t - c_j \rVert_2^2,\qquad z_t^{\mathrm{post}} = 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}} = \lVert\operatorname{sg}[\tilde{z}_t] - z_t^{\mathrm{post}} \rVert_2^2 + \beta\lVert\tilde{z}_t - \operatorname{sg}[z_t^{\mathrm{post}}] \rVert_2^2,

where sg⁡[⋅]\operatorname{sg}[\cdot] is the stop-gradient operator.

The assignment induces the one-hot posterior qϕ(kt∣⋅)=1{kt=ktpost}q_{\phi}(k_t \mid\cdot) = \mathbf{1}\{k_t = k_t^{\mathrm{post}}\}. The prior predicts this assignment from hth_t alone:

Lprior=DKL(qϕ(kt∣⋅) ∥ pθ(kt∣ht))=−log⁡pθ(ktpost∣ht).\mathcal{L}_{\mathrm{prior}} = D_{\mathrm{KL}}\bigl(q_{\phi}(k_t \mid\cdot)\,\|\,p_{\theta}(k_t \mid h_t)\bigr) = -\log p_{\theta}(k_t^{\mathrm{post}} \mid h_t).

The prior-alignment arrow in Figure 1 denotes this cross-entropy. The posterior discovers transition-informed codes from observed futures, while the prior predicts them using only pre-transition information.

Architecture of the profile-conditioned transition and behavior prediction framework

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.

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}}(h_t, z_t, \hat{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}}(h_t, 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\bigl(f_{\theta}^{\mathrm{gate}}(h_t, z_t, \hat{s}_{t+1})\bigr),\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.

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 forms s^t+1r=fθwm(ht,ztr)\hat{s}_{t+1}^{r} = f_{\theta}^{\mathrm{wm}}(h_t, z_t^{r}) for r∈{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}\}} \lVert\hat{s}_{t+1}^{r} - s_{t+1} \rVert_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 behavior at+1∗a_{t+1}^{*},

Lmacro=−log⁡pθ(at+1∗∣qi,st,at).\mathcal{L}_{\mathrm{macro}} = -\log p_{\theta}(a_{t+1}^{*} \mid q_i, s_t, a_t).

The gated decoder is trained through Lmacro\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.

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. Let ξt:t+H\xi_{t:t+H} denote the executable segment over horizon HH. The execution policy is

πωexec(ξt:t+H∣qi,st,at).\pi_{\omega}^{\mathrm{exec}}(\xi_{t:t+H} \mid q_i, s_t, a_t).

Here, qiq_i preserves participant-level conditioning, sts_t grounds execution in the simulator, and ata_t specifies the macro intent.

We train the execution policy with DPO [11]. Under the same context, pairs are ranked by deviation from the corresponding human trajectory: ξ+\xi^{+} is the closer match and ξ−\xi^{-} has substantially larger deviation. The objective is

LDPO=−log⁡σ(β[log⁡πωexec(ξ+∣qi,st,at)πref(ξ+∣qi,st,at)−log⁡πωexec(ξ−∣qi,st,at)πref(ξ−∣qi,st,at)]).(1)\mathcal{L}_{\mathrm{DPO}}=-\log\sigma\left(\beta\left[\log\frac{\pi_{\omega}^{\mathrm{exec}}(\xi^{+}\mid q_i,s_t,a_t)}{\pi_{\mathrm{ref}}(\xi^{+}\mid q_i,s_t,a_t)}-\log\frac{\pi_{\omega}^{\mathrm{exec}}(\xi^{-}\mid q_i,s_t,a_t)}{\pi_{\mathrm{ref}}(\xi^{-}\mid q_i,s_t,a_t)}\right]\right). \tag*{(1)}

where πref\pi_{\mathrm{ref}} is the reference execution policy and β\beta controls preference strength.

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.

Experimental Setup

Simulation Environment. We evaluate our framework on two complementary simulation platforms: AgentSociety [10] and VEET, 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. 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, and move. 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.

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}, 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 of 240 training, 25 validation, and 48 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 reports this subject-level OOD evaluation.

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 is DKL(Pdata(a)∥Ppred(a))D_{\mathrm{KL}}(P_{\mathrm{data}}(a) \mathbin{\|} P_{\mathrm{pred}}(a)), and Dist. Sim. is the macro-action distribution similarity 1−JSD⁡(Pdata,Ppred)/log⁡21 - \operatorname{JSD}(P_{\mathrm{data}}, P_{\mathrm{pred}})/\log2. For Ours on VEET, the second line reports 95% participant-level bootstrap confidence intervals over all 48 test participants (1,000 resamples; seed 2026).

ModelTop-1 ↑\uparrowTop-3 ↑\uparrowMacro KL ↓\downarrowDist. Sim. ↑\uparrowVQ Codes ↑\uparrow
VEET
Random0.1670.5000.9370.666–
Majority0.739–6.8880.855–
Direct SFT / BC0.7540.8900.3610.846–
IRL policy0.6500.8450.4310.758–
Ours0.904 [0.895, 0.913]0.989 [0.987, 0.990]0.0033 [0.0023, 0.0046]0.999 [0.998, 0.999]19
AgentSociety
Random0.2000.6000.0810.869–
Majority0.298–0.8430.504–
Direct SFT0.5250.8260.1580.681–
Ours0.5950.9330.0780.9745

Table 1.

Implementation Details. The trajectory encoder EϕE_{\phi} is a two-layer transformer with hidden dimension 128 and four attention heads. The VQ codebook contains K=128K = 128 entries of dimension 64, updated via exponential moving average with decay 0.99. The profile encoder EψqE_{\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= 3 \times10^{-4}, β1=0.9\beta_1 = 0.9, β2=0.999\beta_2 = 0.999) for 100 epochs with batch size 256. The objective weights are set to λs=1.0\lambda_s = 1.0, λVQ=0.25\lambda_{\mathrm{VQ}} = 0.25, λz=0.1\lambda_z = 0.1, and λa=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; 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. Let PdP_d and PpP_p denote the empirical and predicted macro-action distributions. Macro KL is DKL(Pd∥Pp)D_{\mathrm{KL}}(P_d\|P_p), and macro-distribution similarity is 1−JSD⁡(Pd,Pp)/log⁡21 - \operatorname{JSD}(P_d,P_p)/\log2. 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.

Unseen-Participant Behavioral Prediction

Table 1 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.

Contribution of Transition Supervision

Table 2 separates continuous next-state prediction, discrete transition modeling, and rollout-time prior alignment. The largest point-wise difference is between direct SFT/BC in Table 1 (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.

Table 2: VEET latent representation ablation.

VariantTop-1 ↑\uparrowTop-3 ↑\uparrowMacro KL ↓\downarrow
Simple WM + decoder0.9010.9520.012
VQ-WM0.9050.9810.0078
VQ-WM + prior align.0.9040.9890.0033

Table 2.

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 tele report 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).

Table 3: Descriptive held-out VEET performance by environment-specific questionnaire-profile cluster.

ClusterProfile tagNNTop-1Teleport R.
V-C0Mid profile70.9100.18
V-C1Elevated stress180.9130.13
V-C2Higher stress / BDI80.9050.26
V-C3High hedonic / low BDI150.8910.26

Table 3.

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| \approx0.11–0.19). This result does not prove disentanglement, but it is consistent with the transition modes not being reducible to individual questionnaire dimensions (Figure 3).

Exploratory associations between participant-level prior-code usage frequencies and seven psychometric scales

Figure 3. Exploratory associations between participant-level prior-code usage frequencies and seven psychometric scales. Cells report Spearman ρ\rho; * denotes BH–FDR q<0.05q < 0.05 across 133 tests and † denotes uncorrected p<0.05p < 0.05. No association survives FDR correction. The median pairwise sample size is 297, and codes are ordered by maximum absolute correlation.

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.

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. Figure 2b 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 Table 3 and Table 4. AgentSociety implementation details appear in Appendix § 9.

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. Here, nn denotes the number of held-out macro-action events rather than unique participants. The panels visualize environment-specific outcomes and are not pooled.

Table 4: Descriptive held-out AgentSociety performance by environment-specific questionnaire-profile cluster.

ClusterProfile tagNNTop-1Leisure R.
A-C0Mid profile130.5580.16
A-C1High hedonic / low BDI140.6060.27
A-C2Elevated stress100.6000.25
A-C3Higher stress / BDI110.6130.24

Table 4.

Table 4 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 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 overall move frequency. DPO improves completion and reduces both gaps; on VEET, it also increases coin gain.

Table 5: Micro-control results for SFT/BC and macro-conditioned DPO.

MetricVEET SFT/BCVEET DPOAgentSoc. SFT/BCAgentSoc. DPO
Completion ↑\uparrow0.1680.3320.6640.990
Coin gain ↑\uparrow0.1202.516––
First-act. KL ↓\downarrow2.5561.3821.4680.712
Move. gap ↓\downarrow0.8730.5200.4530.267

Table 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.

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.

Generative AI Usage

Generative AI tools assisted manuscript organization and language editing, LaTeX and figure preparation, and development and debugging of analysis code.

References

AgentSociety

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 an AgentSocProfile. 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 native SocietyAgent instances, we patch each live agent by writing these values into its memory state. Specifically, we initialize hunger_satisfaction, energy_satisfaction, safety_satisfaction, social_satisfaction, and the added pleasure_satisfaction field, 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 the NeedsBlock parameters after agent construction. Questionnaire traits scale the decay rates of hunger, energy, safety, social, and pleasure needs through multiplicative factors αH\alpha_H, αD\alpha_D, αP\alpha_P, αC\alpha_C, αL\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 registered SocialBlock from 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} = \{\mathrm{work}, \mathrm{consume}, \mathrm{rest}, \mathrm{leisure}, \mathrm{move}\}. Internally, these correspond to coarse plan types work, consume, sleep, leisure, and mobility. 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 that rest and leisure correspond 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.

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 Algorithm 1 AgentSociety profile-conditioned simulation

Require: Questionnaire profiles {qi}\{q_i\}, selected subjects SS

  1. for i∈Si \in S do

  2. pi←QuestionnaireToProfile(qi)p_i \leftarrow\text{QuestionnaireToProfile}(q_i)

  3. Create AgentSociety agent aia_i

  4. Inject pip_i into memory needs, emotions, personality, and thought

  5. Patch NeedsBlock decay rates and thresholds

  6. Remove social guidance and disable SocialBlock

  7. Restrict plan steps to {work,consume,rest,leisure,move}\{\text{work}, \text{consume}, \text{rest}, \text{leisure}, \text{move}\}

  8. end for

  9. for each simulation time step tt do

  10. for each agent aia_i do

  11. Observe current simulator state, location, memory, and profile

  12. Compute profile-conditioned macro prior

  13. LLM selects a structured macro action and generates a plan

  14. Execute the corresponding AgentSociety behavior block

  15. Log state, profile, macro action, probabilities, and description

  16. end for

  17. end for

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 vector qiq_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 with ticks_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}\{\text{work}, \text{consume}, \text{rest}, \text{leisure}, \text{move}\}. These event-level macro-action traces form the AgentSociety dataset used for profile-conditioned behavior modeling.

VEET Explanation

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 vector qiq_i, which conditions our models throughout training.

Observation state. At each time step, we convert the raw Unity observation into a flat numerical state vector sts_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 subject ii is associated with a questionnaire-derived vector qiq_i. Depending on the experiment, qiq_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 same qiq_i is concatenated with the environment state or encoded as a 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_coin corresponds to movement that reduces distance to nearby coins, go_to_next_door and go_to_prev_door correspond to room-transition-directed behavior, and teleport corresponds 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), where mtm_t is movement intensity, rtr_t is turn direction, and ztz_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 state sts_t, the subject profile qiq_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}} = \{(s_t, q_i, a_t^{\mathrm{macro}}, s_{t+1}, a_{t+1}^{\mathrm{macro}})\},

whereas the micro controller uses

Dmicro={(st,qi,atmacro,atmicro,st+1)}.\mathcal{D}_{\mathrm{micro}} = \{(s_t, q_i, a_t^{\mathrm{macro}}, a_t^{\mathrm{micro}}, 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.

References

  1. [1]Marwa Abdulhai, Ryan Cheng, Donovan Clay, Tim Althoff, Sergey Levine, and Natasha Jaques. 2025. Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning. arXiv:2511.00222 [cs.CL] https://arxiv.org/abs/2511.00222
  2. [2]Jack Brookes, Matthew Warburton, Mshari Alghadier, Mark Mon-Williams, and Faisal Mushtaq. 2019. Studying human behavior with virtual reality: The Unity Experiment Framework. Behavior Research Methods 52, 2 (April 2019), 455–463. doi:10.3758/s13428-019-01242-0DOI
  3. [3]Yifan Duan, Yihong Tang, Xuefeng Bai, Kehai Chen, Juntao Li, and Min Zhang. 2025. The Power of Personality: A Human Simulation Perspective to Investigate Large Language Model Agents. arXiv:2502.20859 [cs.CL] https://arxiv.org/abs/2502.20859
  4. [4]Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. 2020. Dream to Control: Learning Behaviors by Latent Imagination. arXiv:1912.01603 [cs.LG] https://arxiv.org/abs/1912.01603
  5. [5]Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. 2022. Mastering Atari with Discrete World Models. arXiv:2010.02193 [cs.LG] https://arxiv.org/abs/2010.02193
  6. [6]Danijar Hafner, Jurgen Pasukonis, Jimmy Ba, and Timothy Lillicrap. 2023. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104 (2023).arxiv.org/abs/2301.04104
  7. [7]Samuel R. C. Hewitt, Agnes Norbury, Quentin J. M. Huys, and Tobias U. Hauser. 2025. Day-to-day fluctuations in motivation drive effort-based decision-making. Proceedings of the National Academy of Sciences 122, 12 (2025), e2417964122. arXiv:https://www.pnas.org/doi/pdf/10.1073/pnas.2417964122 doi:10.1073/pnas.2417964122DOI
  8. [8]Michael D. Kozlov and Mark K. Johansen. 2010. Real Behavior in Virtual Environments: Psychology Experiments in a Simple Virtual-Reality Paradigm Using Video Games. Cyberpsychology, Behavior, and Social Networking 13 (2010), 711 – 714. https://api.semanticscholar.org/CorpusID:9400778DOI
  9. [9]Joon Sung Park, Carolyn Q. Zou, Jonne Kamphorst, Niles Egan, Aaron Shaw, Benjamin Mako Hill, Carrie Cai, Meredith Ringel Morris, Percy Liang, Robb Willer, and Michael S. Bernstein. 2026. LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals. arXiv:2411.10109 [cs.AI] https://arxiv.org/abs/2411.10109
  10. [10]Jinghua Piao, Yuwei Yan, Jun Zhang, Nian Li, Junbo Yan, Xiaochong Lan, Zhihong Lu, Zhiheng Zheng, Jing Yi Wang, Di Zhou, et al. 2025. Agentso-ciety: Large-scale simulation of llm-driven generative agents advances understanding of human behaviors and society. (2025).arxiv.org/abs/2502.08691
  11. [11]Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems 36 (2023), 53728–53741.arxiv.org/abs/2305.18290
  12. [12]Mingjun Ren and Wentao Xu. 2025. The Impact of Big Five Personality Traits on AI Agent Decision-Making in Public Spaces: A Social Simulation Study. arXiv:2503.15497 [cs.HC] https://arxiv.org/abs/2503.15497
  13. [13]Yiding Wang, Yuanxuan Chen, Fangwei Zhong, Long Ma, and Yizhou Wang. 2025. Simulating Human-like Daily Activities with Desire-driven Autonomy. arXiv:2412.06435 [cs.AI] https://arxiv.org/abs/2412.06435

Paper details

vlnc:W_jcmygm53
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Generative AI tools

View statement
Looking for related papers…
Contents