Cross-Centaur: Modeling Individual Decision-Making Across Psychological Tasks with Language Models

Hikaru Asano; Hidezo Suganuma; Ryutaro Mori; Yukino Baba

cs.CLOct 16, 2026

Cross-Centaur: Modeling Individual Decision-Making Across Psychological Tasks with Language Models

Authors: Hikaru Asano, Hidezo Suganuma, Ryutaro Mori, Yukino Baba

Organizations: The University of Tokyo · RIKEN Center for Brain Science

Abstract

Large language models (LLMs) are increasingly used as computational models of human behavior. Recent models such as Centaur have advanced this goal by fine-tuning LLMs on trial-by-trial decisions from psychological experiments. However, existing large-scale behavioral datasets do not link the same participants across different tasks, so these models learn population-level regularities but are not trained to model individual-level regularities across tasks. We introduce Cross-Psych-101, a dataset of 591 participants and 261{,}551 valid decisions across 12 Bandit, Tradeoff, Structure, and Social tasks. We train Cross-Centaur to predict a participant's target-task actions from a target-task description and the source-task description paired with the same participant's source-task behavior. Cross-Centaur outperforms off-the-shelf LLMs and Centaur on held-out participants, both within and across tasks. Probing analyses suggest that its hidden states encode participant-level behavioral tendencies that transfer across tasks. These results highlight the potential of cross-task individual-level data for predicting human decision-making with language models. Cross-Pysch-101 is available at https://huggingface.co/datasets/HikaruAsano/cross-psych-101.

Figures & tables

Figure 1: Comparison of Psych-101 and Centaur with Cross-Psych-101 and Cross-Centaur. Psych-101 observes participants within individual tasks and supports within-task behavioral prediction, whereas Cross-Psych-101 follows the same participants across tasks, enabling Cross-Centaur to use source-task behavior to predict the same individual’s behavior in a different target task.

Figure 2: Overview of the Cross-Psych-101 dataset collection. The suite comprises twelve tasks spanning four domains (Bandit, Tradeoff, Structure, and Social). Each participant completes tasks from multiple domains, enabling analysis of individual behavioral signatures within and across domains.

Figure 3: Summary of the data collection protocol. (1) Participants were recruited on Prolific with eligibility screening. (2) Each was randomly assigned two of four task categories and completed three tasks per category. (3) Quality control included reCAPTCHA, attention checks, response-time and browser-focus filters, and an opt-out for public release. The released dataset contains 261,551 valid decisions from 591 participants.

Figure 4: Validation analyses for four tasks, with one task selected from each category. Each panel summarizes participant-level performance or preferences in a different task. (A) Mean regret in the Two-Armed Bandit task. Shaded areas indicate standard errors. (B) Distribution of the proportion of trials in which participants chose the riskier lottery in the One-Shot Gamble task. (C) Mean absolute error in the Estimation task, shown separately for each target function. Shaded areas indicate standard errors. (D) Distribution of Social Value Orientation (SVO) scores. For the remaining eight tasks, see Appendix Fig. 7 .

Model	Centaur	Cross-Centaur Same-Task				Cross-Centaur Cross-Task
	Overall	Bandit	Tradeoff	Structure	Social	Bandit		Tradeoff		Structure		Social
						Same cat.	Other cat.	Same cat.	Other cat.	Same cat.	Other cat.	Same cat.	Other cat.
Qwen3-8B Base	0.680	0.928	1.031	0.838	1.053	0.902	0.865	0.913	0.754	0.836	0.834	1.160	1.208
Qwen3-8B w/ reasoning	1.009	1.681	1.522	1.421	1.552	1.531	1.159	1.267	0.959	1.265	1.294	1.620	1.563
Qwen3-8B w/ hints	0.680	1.996	1.476	1.517	1.798	1.781	1.209	1.254	1.019	1.371	1.343	1.712	1.538
Qwen3-14B Base	0.672	0.916	1.025	0.847	1.022	0.907	0.883	0.898	0.735	0.828	0.819	1.136	1.181

Table 1: Performance breakdown across Cross-Centaur evaluation settings, measured by negative log likelihood (lower is better). “Same cat.” and “Other cat.” denote whether the source task belongs to the same psychological category as the target task. For Centaur and Cross-Centaur, we report the average over three training seeds.

2-Armed Bandit

Restless Bandit

Correlated Bandit

One-Shot Gamble

Temporal Discounting

Shuffled Card

Classification

Estimation

Odd One Out

Public Goods

Social Value

Ultimatum Game

Qwen3-8B

0.937 / 0.607

1.201 / 1.140

3.602 / 3.370

0.841 / 0.709

1.129 / 0.805

1.904 / 1.953

1.030 / 1.018

1.230 / 1.413

1.788 / 1.397

1.318 / 1.341

1.908 / 1.714

5.031 / 3.757

Qwen3-14B

1.290 / 0.616

1.416 / 1.362

4.235 / 4.860

0.970 / 0.812

1.638 / 0.617

2.162 / 2.325

1.526 / 1.727

1.265 / 1.498

1.324 / 1.217

1.390 / 1.359

3.164 / 2.202

6.525 / 4.747

Centaur

0.300 / 0.315

0.804 / 0.815

1.493 / 1.819

0.411 / 0.505

0.274 / 0.294

1.358 / 1.633

0.691 / 0.684

0.928 / 0.996

0.798 / 0.814

0.808 / 0.870

1.613 / 1.618

1.583 / 1.830

Cross-Centaur

0.294 / 0.313

0.786 / 0.808

1.462 / 1.780

0.411 / 0.520

0.269 / 0.292

1.365 / 1.630

0.681 / 0.676

0.925 / 0.987

0.803 / 0.821

0.810 / 0.865

1.629 / 1.649

1.554 / 1.777

Table 2: Task-level negative log likelihood on Cross-Psych-101. Lower values are better. Each cell reports participant-held-out / task-held-out performance for the corresponding target task.

	Bandit	Tradeoff	Structure	Social
Qwen3-8B / Shuffled	1.731	1.463	1.393	2.115
Qwen3-8B / Target	1.708	1.395	1.318	1.630
Qwen3-8B / Orig	1.689	1.336	1.403	1.756
Centaur / Shuffled	0.964	1.756	1.093	1.697
Centaur / Target	0.969	1.899	1.125	1.635
Centaur / Orig	0.932	1.580	1.081	1.612

Table 3: Performance breakdown (negative log-likelihood, lower is better) by task family. Abbreviations: Orig.=Original prompt, Shuffled=Shuffled source, Target=Target only.

Model	Same-Task				Cross-Task
	Bandit	Tradeoff	Structure	Social	Bandit		Tradeoff		Structure		Social
					Same cat.	Other cat.	Same cat.	Other cat.	Same cat.	Other cat.	Same cat.	Other cat.
Qwen3-8B	0.724	0.759	0.567	0.603	0.583	0.567	0.531	0.438	0.533	0.533	0.596	0.574
Centaur	0.738	0.759	0.600	0.603	0.600	0.556	0.542	0.438	0.533	0.533	0.596	0.569
Cross-Centaur	0.734	0.760	0.633	0.603	0.634	0.609	0.749	0.477	0.567	0.632	0.608	0.553

Table 4: Probe accuracy breakdown by psychological category on Cross-Psych-101. Higher values are better. “Same cat.” and “Other cat.” denote whether the source task belongs to the same psychological category as the target task.

Figure 5: Cross-Centaur hidden states visualized by shuffled-card-task risk-preference score. Each point represents a participant, colored by their observed risk preference (yellow = more risk-seeking, blue = more risk-averse). The horizontal axis shows the hidden state’s projection onto a linear probe for this score, so more risk-seeking participants appear further right; the vertical axis is orthogonal and just spreads points in two dimensions. Rows vary the source task (top: shuffled card, same task; bottom: restless bandit, cross task), and columns compare inputs with only source-task behavior (left) versus those also including the shuffled-card target-task description (right). Adding the target description clarifies the left-to-right risk preference gradient, especially for the cross-task (bottom) row.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Screenshots of the experimental interfaces for the 12 tasks. Each task is classified under one of the following four categories: Bandit, Tradeoff, Structure, or Social.

Figure 7: Supplementary validation analyses across eight tasks not discussed in the main text. Panels (A) and (B) show mean regret in the Restless and Correlated bandit tasks, respectively. Panel (C) shows the distribution of the proportion of trials in which participants chose the delayed option in the Temporal Discounting task. Panel (D) shows the distribution of mean session length, a proxy for risk preference, in the Shuffled Card task. Panel (E) shows mean error rate in the Classification task, separately for each classification rule. Shaded areas indicate standard errors. Panel (F) shows the margin-weighted match rate in the Odd-One-Out task, which can be interpreted as the consistency between participants’ subjective similarity judgments and distances derived from pretrained word-embedding vectors. Panel (G) shows mean contributions in the conditional Public Goods Game task. Panel (H) shows a scatterplot of each participant’s proposal and minimum acceptance threshold in the Ultimatum Game task.

Figure 8: Task co-occurrence counts across participants. Each cell shows the number of participants who completed both tasks in the corresponding unordered task pair.

Task Name	Category	Incentivized?	Reference	Summary
Two-armed bandit	Bandit	Yes	Gershman (2018)	Choose between a safe option (always yielding zero points) and a risky option (yielding a positive or negative reward). 10 trials $\times$ 8 sessions; the risky option’s reward distribution varies across sessions.
Restless bandit	Bandit	Yes	Daw et al. (2006)	Choose one of four options whose expected rewards drift over time. 50 trials $\times$ 2 sessions.
Correlated bandit	Bandit	Yes	Wu et al. (2018)	Choose one of thirty options arranged in a row, where neighboring options yield similar rewards. 10 sessions in total (5 long sessions of 10 trials and 5 short sessions of 5 trials); reward distributions and scales vary across sessions.
One-shot gamble	Tradeoff	Yes	Krueger et al. (2024)	Choose between two lotteries, each specified by probabilities and payoffs, for 20 trials. One chosen lottery is resolved at the end of the task.
Temporal discounting	Tradeoff	No	Ruggeri et al. (2022)	Choose between an immediate (sooner) and a delayed (later) monetary option across varying amounts and delays. 3 trials $\times$ 4 sessions, covering both gain and loss conditions.
Shuffled card task	Tradeoff	Yes	Frey et al. (2017)	Repeatedly draw a card from a deck of 32 (containing gain and bomb cards) or stop. Drawing a bomb card forfeits the session’s accumulated points and ends the session. 35 sessions, varying in gain amount, loss amount, and number of bomb cards.

Appendix

Table 5: List of the 12 tasks used in the experiment. This table details the category, the presence of financial incentives, the reference, and the specific task description for each task. Note that due to time constraints, the number of trials and sessions has generally been reduced compared to the original tasks.

Setting	Value
Base model	Qwen3-8B
Training framework	Unsloth
Hardware	Eight NVIDIA GH200 Grace Hopper Superchip GPUs
Trainable parameters	LoRA adapter weights only
LoRA configuration	Rank $r=16$ , $\alpha=32$ , dropout $0.05$ , no bias terms
LoRA target modules	q_proj , k_proj , v_proj , o_proj , gate_proj , up_proj , down_proj

Appendix

Table 6: Fine-tuning hyperparameters for the Qwen3-8B-based Cross-Centaur models.

Setting	Value
Model variants	Qwen3-8B, Centaur, and Cross-Centaur
Representation input	User-side prompt only; assistant target transcript excluded
Hidden-state layer	Final transformer layer
Pooling	Mean pooling over non-padding tokens
Model inference precision	4-bit model loading with bfloat16 computation
Probe scope	Separate probe for each task

Appendix

Table 7: Hyperparameters for extracting model representations and training behavioral tendency probes.

Task	Score Name	Type	Definition	Interpretation of Larger Values
Two-armed bandit	regret	performance	Negative mean regret, where regret is best reward minus realized reward.	Better choices.
Restless bandit	regret	performance	Negative mean regret, where regret is the maximum mean reward on each trial minus the realized reward.	Better choices.
Correlated bandit	regret	performance	Negative mean regret, where regret is rewardScale $\times\max(\mu)$ minus the realized reward.	Better choices.
One-shot gamble	risk preference	risk preference	Fraction of the 20 trials in which the participant chose the higher-variance option.	More risk-seeking choices.
Shuffled card tas	risk preference	risk preference	Number of sessions ending by drawing a bomb, divided by 35.	More sessions in which the participant continues drawing until a bomb.
Shuffled card task	risk preference	risk preference	Mean stopping trial among voluntarily stopped sessions, divided by 35.	Later stopping and stronger tendency to keep drawing.

Appendix

Table 8: Participant-level behavioral tendency scores used as targets for probing and as color variables for representation visualization.

Figure 9: Comparison between human and Cross-Centaur regret in the Two-armed bandit task. The scatterplot shows participant-level regret values for humans and Cross-Centaur predictions, with a correlation of $r=0.840$ . The distribution plot shows that human regret is more broadly distributed than Cross-Centaur regret.

Model	Cross-Centaur Same-Task				Cross-Centaur Cross-Task
Bandit	Tradeoff	Structure	Social	Bandit		Tradeoff		Structure		Social
				Same cat.	Other cat.	Same cat.	Other cat.	Same cat.	Other cat.	Same cat.	Other cat.
Qwen3-8B-prompt Shuffled source	2.076	1.822	1.431	2.569	1.837	1.281	1.488	1.080	1.368	1.379	2.031	1.746
Qwen3-8B-prompt Target only	1.907	1.709	1.374	1.638	1.751	1.465	1.432	1.044	1.303	1.277	1.628	1.624
Qwen3-8B Original	2.007	1.587	1.458	1.744	1.798	1.262	1.350	1.071	1.371	1.379	1.783	1.741
centaur prompt Shuffled source	1.071	2.218	1.146	1.709	0.987	0.834	1.761	1.291	1.078	1.055	1.691	1.691

Appendix

Table 9: Performance breakdown across Cross-Centaur evaluation settings. Lower values are better. “Same cat.” and “Other cat.” denote whether the source task belongs to the same psychological category as the target task.

Model	Centaur	Cross-Centaur Same-Task				Cross-Centaur Cross-Task
	Overall	Bandit	Tradeoff	Structure	Social	Bandit		Tradeoff		Structure		Social
						Same cat.	Other cat.	Same cat.	Other cat.	Same cat.	Other cat.	Same cat.	Other cat.
Centaur	$0.652\pm 0.002$	$0.861\pm 0.003$	$1.069\pm 0.019$	$0.856\pm 0.001$	$1.043\pm 0.002$	$0.847\pm 0.003$	$0.828\pm 0.002$	$0.950\pm 0.011$	$0.802\pm 0.000$	$0.865\pm 0.000$	$0.875\pm 0.001$	$1.171\pm 0.001$	$1.204\pm 0.001$
Cross-Centaur	$0.646\pm 0.002$	$0.812\pm 0.002$	$0.987\pm 0.029$	$0.817\pm 0.001$	$1.009\pm 0.000$	$0.799\pm 0.002$	$0.787\pm 0.003$	$0.872\pm 0.017$	$0.720\pm 0.001$	$0.806\pm 0.001$	$0.809\pm 0.001$	$1.127\pm 0.002$	$1.160\pm 0.001$

Appendix

Table 10: Variability of Centaur and Cross-Centaur across training seeds. Each cell reports the mean

\pm

standard deviation of negative log likelihood over the three training seeds used in Table 1 (lower is better). “Same cat.” and “Other cat.” denote whether the source task belongs to the same psychological category as the target task.

		Cross-Psych-101
Baseline	Psych-101	Same-Task	Cross-Task
Qwen3-8B Base	1.350 [1.232, 1.473]	1.254 [1.051, 1.478]	0.899 [0.666, 1.199]
Centaur	0.008 [0.003, 0.014]	0.110 [0.070, 0.170]	0.150 [0.084, 0.241]

Appendix

Table 11: Bootstrap comparison against Cross-Centaur. Each cell reports

\Delta\mathrm{NLL}=\mathrm{NLL}_{\text{baseline}}-\mathrm{NLL}_{\text{Cross-Centaur}}

with 95% confidence intervals from 10,000 participant-level bootstrap resamples of the test data; positive values favor Cross-Centaur.

Model	Same-Task				Cross-Task
	Bandit	Tradeoff	Structure	Social	Bandit	Tradeoff	Structure	Social
Llama-3.2-1B	1.193 / 1.040	1.865 / 1.590	1.197 / 1.095	1.574 / 1.531	1.057 / 0.908	1.420 / 1.149	1.192 / 1.030	1.800 / 1.667
Llama-3.1-8B	0.986 / 0.938	1.641 / 1.252	1.081 / 1.000	1.563 / 1.373	0.862 / 0.805	1.232 / 0.834	1.002 / 0.883	1.714 / 1.462
Qwen3-1.7B	0.906 / 0.861	1.155 / 1.028	0.994 / 0.871	1.128 / 1.082	0.917 / 0.853	1.028 / 0.868	1.040 / 0.869	1.330 / 1.222
Qwen3-8B	0.854 / 0.792	1.058 / 0.910	0.856 / 0.795	1.042 / 0.982	0.831 / 0.765	0.873 / 0.729	0.869 / 0.772	1.184 / 1.107

Appendix

Table 12: Performance across model families and sizes on the participant-held-out split. Each cell reports Centaur / Cross-Centaur negative log likelihood for the corresponding task family; lower values are better, and the better of the two is in bold.

Panel of Figure 10 — Figure 10: Prompt template for Cross-Centaur: The formatting for user and assistant messages during supervised fine-tuning, with color highlighting for key sections.

Human Behavior Modeling

See all 27 papers

Post-training makes large language models less human-like
May 8, 2026Marcel Binz, Elif Akata, Abdullah Almaatouq +76LLM Evaluation LLM Alignment
Small Foundation Models of Human Cognition and Behaviour
Aug 5, 2026Nick Oh, Fernand GobetComputational Cognitive Modeling Human Behavior Modeling
socius labs · London School of Economics and Political Science
ActTraitBench: Quantifying the Knowledge-Decision Gap in Large Language Models via Human-Grounded Behavioral Validation
May 28, 2026Yutong Yang, Chenxi Miao, Weikang Li +1LLM Evaluation Personality Modeling in Language Models
Peking University · Baidu Inc

May 8, 2026J/K move · Enter open · S save

Post-training makes large language models less human-like

Marcel Binz, Elif Akata, Abdullah Almaatouq, Mohammed Alsobay+75

LLM Evaluation LLM Alignment Human Behavior Prediction Language Model Post-Training Persona-Conditioned Generation Human Behavior Modeling

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly used as surrogates for human participants, but it remains unclear which models best capture human behavior and why. To address this, we introduce Psych-201, a novel dataset that enables us to measure behavioral alignment at scale. We find that post-training -- the stage that turns base models into useful assistants -- consistently reduces alignment with human behavior across model families, sizes, and objectives. Moreover, this misalignment widens in newer model generations even as base models continue to improve. Finally, we find that persona-induction -- a popular technique for eliciting human-like behavior by conditioning models on participant-specific information -- does not improve predictions at the level of individuals. Taken together, our results suggest that the very processes that are currently employed to turn LLMs into useful assistants also make them less accurate models of human behavior.

Abstract

Human Behavior Modeling

Explore similar work

Post-training makes large language models less human-like

Small Foundation Models of Human Cognition and Behaviour

ActTraitBench: Quantifying the Knowledge-Decision Gap in Large Language Models via Human-Grounded Behavioral Validation