Skip to main content
Back

Cross-Centaur Predicts Your Choices in Tasks It Has Never Seen You Do

A new dataset tracks the same people across 12 psychology tasks, letting a fine-tuned LLM predict individual decisions rather than population averages.

  • Cross-Centaur, accepted as an EMNLP 2026 oral, predicts individual decisions across psychology tasks using fine-tuned LLMs.
  • Introduces Cross-Psych-101: 591 participants, 261,551 decisions, 12 tasks across Bandit, Tradeoff, Structure, Social domains.
  • Same participants complete multiple tasks, enabling individual-level transfer that Psych-101 could not support.
  • Beats off-the-shelf LLMs and the original Centaur on held-out participants, both within and across tasks.
  • Linear probes on hidden states recover participant-level traits like risk preference that transfer between tasks.
  • Dataset is public on Hugging Face; model weights are not currently released with it.

The earlier Centaur project fine-tuned a language model on trial-level records from psychology experiments. Its authors reported stronger predictions for held-out participants, meaning people excluded from training, than established cognitive models. The model also generalized to new cover stories, modified task structures, and different domains. An EMNLP 2026 paper, Cross-Centaur, extends that approach to person-specific prediction across tasks.

Psych-101 gave Centaur more than 10 million decisions from over 60,000 participants across 160 experiments. Its anonymized records lack links between the same participant’s sessions in different experiments. Centaur can therefore learn population patterns such as typical risk tolerance or temporal discounting, while stable differences between individuals remain unavailable. Cross-Centaur uses linked histories to estimate how one person’s behavior may transfer from one task to another.

Identity Survives the Task Switch

The authors created Cross-Psych-101, a dataset containing 261,551 valid decisions from 591 participants across 12 tasks. The tasks span four behavioral domains:

Domain Tasks Behavior measured
Bandit 3 Exploration and exploitation
Tradeoff 3 Risk preference and temporal discounting
Structure 3 Rule learning and similarity judgments
Social 3 Ultimatum games, public goods, and social value orientation

Participants recruited through Prolific were randomly assigned two domains and completed all three tasks in each, producing six linked task histories per person. Quality controls included reCAPTCHA, attention checks, and response-time filters. This repeated-measures design gives the model several observations of each participant while reserving other tasks as prediction targets. The dataset is available on