Can language models predict individual decisions across psychological tasks?

Population patterns are not individual predictions

Most large behavioral datasets used to train language models of human choice record each participant on a single task. A model trained on such data can learn how people behave on average in many tasks, but it never sees the same person in two different settings, so it cannot learn what makes one individual differ from another. The paper states this directly: existing large-scale behavioral datasets do not link the same participants across tasks, so models learn population-level regularities but are not trained to model individual-level regularities across tasks.

That distinction matters for the question in the title. Predicting that a typical participant will prefer a delayed reward is a different problem from predicting that this particular participant, who explored heavily in a bandit game, will make a specific choice in a temporal discounting trial. The second requires evidence that ties behavior to a person rather than to a task.

One dataset follows the same people across tasks

Cross-Psych-101 was built to supply that missing link. It contains 591 participants and 261,551 valid decisions across 12 tasks grouped into four categories: Bandit, Tradeoff, Structure, and Social. The same participants complete multiple tasks, which is what allows a model to be trained on one person's behavior in one task and evaluated on that same person's behavior in another.

The collection protocol, described in the author contributions, assigned participants to task categories randomly with balanced quotas and used performance-based bonuses, alongside quality controls such as comprehension and attention checks and response-time logging. The dataset is therefore a designed cross-task panel rather than a re-labeling of existing single-task data.

Source behavior is the input, target choices are the output

Cross-Centaur is fine-tuned on a prompt with three parts: a description of the source task, the same participant's trial-by-trial behavior in that source task, and a description of the target task. The model must then output the target task's transcript, meaning the participant's predicted choices. The training objective is an action-only language-modeling loss applied to tokens inside the marked action spans, so the model is optimized on the choice tokens rather than on the surrounding text.

The base model is Qwen3-8B, adapted with LoRA weights (rank 16, alpha 32) for one epoch at a learning rate of 2×10⁻⁶ on eight NVIDIA GH200 GPUs. Because only adapter weights are trained, the comparison against Centaur and off-the-shelf Qwen3-8B isolates the effect of the cross-task training data and objective rather than a difference in base model capacity.

Cross-Centaur lowers prediction error against both baselines

Performance is reported as negative log likelihood (NLL) of the participant's actual choices, where lower is better. In the seed-variability table, Centaur's overall NLL is 0.652±0.002 and Cross-Centaur's is 0.646±0.002. The gap widens in the cross-task settings: for a target in the same category as the source, Centaur scores 0.861±0.003 and Cross-Centaur 0.812±0.002; when the source comes from another category, Centaur scores 1.069±0.019 and Cross-Centaur 0.987±0.029.

The bootstrap comparison in Table 11 reports ΔNLL = NLL_baseline − NLL_Cross-Centaur with 95% confidence intervals from 10,000 participant-level resamples, and positive values favor Cross-Centaur. Against Qwen3-8B Base the gains are 1.350 [1.232, 1.473] on Psych-101, 1.254 [1.051, 1.478] in the same-task setting, and 0.899 [0.666, 1.199] in the cross-task setting. Against Centaur the corresponding gains are much smaller but still positive: 0.008 [0.003, 0.014], 0.110 [0.070, 0.170], and 0.150 [0.084, 0.241]. All intervals exclude zero.

The gain depends on whose behavior is shown

Ablations in Table 9 separate the contribution of participant-specific source behavior from the rest of the prompt. When the source transcript is shuffled, Cross-Centaur's NLL rises: for example, in the same-category cross-task Bandit setting it moves from 0.802 with the original source to 0.877 with a shuffled source, and in the other-category Social setting from 1.094 to 1.130. Removing the source entirely and giving only the target description produces intermediate values, such as 0.913 for same-category cross-task Bandit.

The ordering matters for interpretation. If the model were only reading the target task description, shuffling the source would not hurt. The fact that the original, correctly paired source behavior gives the lowest error supports the paper's claim that the improvement comes from using that participant's own earlier choices, not merely from seeing more text.

Hidden states carry person-level tendencies

The probing analyses ask whether the model's internal representations contain information about a participant's behavioral tendencies. Hidden states were extracted from the prompt before the target transcript was generated, so the probe could only use information available before any target behavior existed. Separate single-hidden-layer MLP probes were trained for task-specific tendency scores, with each numeric score binarized at the median of its target column and the probe predicting whether a participant scored above that median.

The paper reports that these probes work in both same-task and cross-task inputs, and that Cross-Centaur's hidden states encode participant-level behavioral tendencies that transfer across tasks. This is presented as suggestive evidence about representation, not as a demonstration that the model uses those tendencies causally. A related visualization projects representations onto a fitted probe direction for the shuffled-card-task risk-preference score, with only the horizontal coordinate informative about the decoded tendency.

The advantage narrows when the task itself is unseen

The paper is explicit that the gains are more modest in the task-held-out setting. The cross-task results also split by whether the source task belongs to the same psychological category as the target: same-category transfers are consistently easier than other-category transfers, as the NLL values above show. The robustness check across backbones (Llama-3.2-1B, Llama-1.1-8B, Qwen3-1.7B, Qwen3-8B) finds Cross-Centaur better than Centaur in both same-task and cross-task settings for every backbone, and larger models perform better within each family, indicating the objective is not specific to one model family.

Scope limits the claim further. Cross-Psych-101 covers controlled psychological tasks in four categories and does not capture the full range of decisions people make in everyday, institutional, or socially embedded settings. Data were collected in English from English-fluent participants, which may limit generalization to other languages and cultural contexts where the same task descriptions could be interpreted differently.

Aggregate fit can hide narrower individual spread

One result shows why a single correlation is not enough to judge individual-level fidelity. In the two-armed bandit task, participant-level regret for humans and for Cross-Centaur predictions correlates at r = 0.840, which looks strong. Yet the accompanying distribution plot shows that human regret is more broadly distributed than Cross-Centaur regret.

Read together, the two views say different things: the model ranks participants similarly to how humans rank, but it compresses the spread of outcomes, producing fewer extreme regret values than people do. For anyone using such a model to simulate individuals rather than averages, that compression is the more consequential signal, and it is the kind of check that a headline correlation would miss.

The dataset, not the score, is the reusable asset

The most decision-relevant comparison in this paper is not Cross-Centaur versus Qwen3-8B, where the NLL gap is large, but Cross-Centaur versus Centaur, where the gap is small and consistent: 0.008 overall, 0.110 same-task, and 0.150 cross-task, with confidence intervals excluding zero in every case. That small but reliable margin is what a cross-task individual-level objective buys over a population-level model trained on the same kind of tasks.

The durable contribution is therefore the linked design: 591 participants and 261,551 valid decisions that follow the same people across 12 tasks. The paper frames Cross-Psych-101 and Cross-Centaur as a foundation for future work toward unified models of individual human decision-making. The practical priority that follows is to judge such models on held-out participants and on transfers to unseen tasks and categories, and to check the spread of predicted behavior, not only its average fit.

Sources used in this answer