# A Foundation Model to Measure and Predict Human Behaviour

AI.info

> Chiron couples a jointly fine-tuned model with 288,307 person records across 78 social-science experiments, offering a concrete tool for measuring individuals and matching population statistics…

**arXiv**: 2611.24830

**Published**: 2026-11-18

**Authors**: Haoran Ye, Tianrui Huo, Hangzhen Li, Xin Zhang, Guojie Song

## AI summary

## Overview

- **Research area:** Natural Language Processing / computational social science — using large language models for psychological measurement, individual behaviour prediction, and population simulation.
- **Technical level:** Advanced (assumes familiarity with LLM fine-tuning, psychometrics, and experimental statistics, though the motivation is explained plainly).
- **Scope:** The paper introduces Chiron, a single fine-tuned language model trained to score 109 psychological constructs, predict how specific individuals will answer unseen questions, and simulate human experimental cohorts whose response distributions match published human statistics.

## What This Paper Is About

General-purpose LLMs can produce fluent-sounding psychological scores, behavioural predictions, and simulated survey respondents, but the paper argues they systematically misrepresent people: they inflate socially desirable traits, compress individual differences, and flatten simulated populations so that experimental effects look far larger than they really are. The authors' goal is to move past prompting chatbots and build a purpose-trained foundation model that is calibrated against real psychometric and experimental data across all three tasks at once.

## Key Contributions

1. **A decoupled measurement head.** Rather than asking the model to write out a number, Chiron attaches a "centred cosine readout" to the internal representation of a person's records, producing scores for 109 psychological constructs directly and avoiding biases that come from text generation.
2. **Joint psychometric and behavioural supervision.** Measurement loss on the 109 constructs shapes the same person representation that the model uses to predict withheld answers, tying interpretable traits to out-of-sample behavioural prediction.
3. **Constrained population synthesis.** Because raw participant-level experimental data are rarely public, the authors reconstruct virtual respondent cohorts whose means, standard deviations, and treatment effects match published summaries, then train the model to generate diverse cohorts per condition.
4. **A large multi-source corpus and benchmark.** Training and evaluation span 36 behavioural and psychological domains and 78 social science experiments, with 288,307 person records, 1,411,750 atomic behavioural decisions, and 99,953 psychometric assessments over 109 psychological constructs.

## Main Findings

- **Better agreement with self-reports:** Chiron's construct scores correlate with human self-report at a mean Pearson r = 0.390 across five datasets and nine constructs, versus 0.320–0.353 for frontier LLMs.
- **Recovers known construct structure:** Inferred scores reproduce literature-reported inter-construct correlations at r = 0.806–0.819 across 150 construct pairs.
- **Fewer psychometric biases:** The paper reports mitigation of score compression, sensitivity to irrelevant context, and sensitivity to evidence order, and a monotonic response to manipulations of construct-relevant evidence across all 109 constructs.
- **More accurate individual prediction:** Mean prediction accuracy of 0.504 across 29 behavioural domains, versus 0.393–0.418 for frontier LLMs (up to 28.4% relative improvement). Gains hold for unseen individuals and for domains excluded from prediction training.
- **Calibrated population simulation:** Across 15 held-out experimental studies, 90.5% agreement on effect direction and 82.2–94.3% lower effect-size estimation error, with a Cohen's d MAE of 0.224 versus 1.254–3.895 for frontier LLMs. Figure 1d reports this MAE comparison over 25 eligible effects from eight studies.
- **Synthetic cohorts match published statistics:** On the 63 training studies, synthesised populations reach Cohen's d MAE = 0.012, r = 0.999, and distribution NMAE = 0.010.
- **Diagnosis of why LLMs fail:** The authors attribute the failures to training objectives that reward helpful, socially desirable, modal responses — noting evidence that RLHF reduces output entropy and collapses conceptual diversity — which is mismatched to psychometric validity and distributional fidelity.
- **Data-efficiency mechanisms:** Self-supervised masked observation prediction converts unlabelled personal blogs into behavioural training instances, and calibrated narratives (25,000 generated narratives, 500,000 judged pairs, 15,250 quality-filtered narratives scored by a Bradley–Terry model) extend text supervision to constructs lacking human text-score pairs.

## Methodology in Plain English

The starting point is an off-the-shelf LLM (Qwen3.6-35B-A3B), adapted with LoRA. Everything about a person — questionnaire answers, essays, social-media posts — is placed in a marked block, and one internal vector at the end of that block becomes the person's "representation." A small learned head turns that vector into 109 trait scores using per-construct directions, scales, and offsets, after subtracting the average representation so generic text properties do not dominate. Three training signals are optimised together: (1) match the head's scores to real self-reports, to scores computed from withheld questionnaire items, and to synthetic questionnaire and narrative targets; (2) predict a person's actual withheld answers or choices from their visible records, including questions built from held-out first-person observations in unlabelled blogs; (3) generate respondent answers whose aggregate statistics reproduce published experimental means, SDs, and effect sizes. Because each experimental condition is paired with many different target respondents under the same prompt, the model is pushed away from producing one modal answer. The combined loss is the prediction loss plus weighted measurement and population terms (λ_m = 8, λ_p = 1), trained for two epochs.

## Why This Matters

**Research impact.** The paper reframes LLM-based behavioural science as a calibration problem rather than a prompting problem, and provides an open model, data, and code release. Its evaluation targets psychometric validity, out-of-sample prediction, and distributional accuracy together, which are usually studied separately.

**Real-world applications.**

- Pretesting experimental designs, framings, and stimuli in silico before running costly human-subject studies.
- Non-invasive psychological assessment from naturalistic text such as essays or social-media histories, where fixed questionnaires are impractical.
- Estimating likely response distributions for subgroups when only aggregate published statistics are available.
- Rapid, interpretable trait profiling for research recruitment, personalisation research, or organisational surveys.

**Industry relevance.** Any organisation that wants to substitute or supplement survey panels — market research, user research, HR and people analytics, health-adjacent text analytics — has a stake in whether simulated respondents have realistic variance. The paper's central warning is that uncalibrated LLMs exaggerate effect sizes and produce false positives on null results, which is directly relevant to anyone making decisions from simulated data.

## Future Directions

- **Closing the aggregate-data gap:** population training relies on published summary statistics rather than raw participant responses; how much accuracy is lost relative to true individual-level supervision is an open question.
- **Validation beyond the current coverage:** the model covers 109 constructs and 78 experiments drawn from specific journals and instruments; transfer to new construct families, languages, and cultural contexts remains to be tested.
- **Guardrails for deployment:** the paper frames Chiron as an instrument for pretesting and measurement; questions of consent, misuse for inferring traits about individuals without their knowledge, and appropriate reporting standards are not resolved here.
- **Separating the three objectives:** the joint training recipe uses fixed weights (λ_m = 8, λ_p = 1); whether measurement, prediction, and population alignment trade off against each other, and how to tune that balance, is left for future work.

## Target Audience

Researchers in computational social science and psychometrics evaluating LLMs as measurement instruments; NLP researchers interested in preference collapse, distributional alignment, and multi-task fine-tuning; and applied scientists in survey research, behavioural economics, or user research who want calibrated simulated respondents rather than plausible-sounding ones. Readers without an LLM background can follow the motivation and results but will need some psychometrics (construct validity, Cohen's d, scale scoring) to interpret the evaluation.

## Authors’ abstract

Large language models (LLMs) are increasingly used to infer psychological attributes, predict individual behaviour, and simulate human populations. However, despite generating plausible responses, they often misrepresent individuals and distort population statistics. Here we introduce Chiron, a foundation model to measure and predict human behaviour. We jointly fine-tune an LLM to represent psychological constructs, predict individual behaviour, and align population response…

[The full abstract on arXiv](https://arxiv.org/abs/2611.24830)

[Read the original paper](https://arxiv.org/abs/2611.24830)

---

Canonical: https://ai.info/paper/a-foundation-model-to-measure-and-predict-human-behaviour-2611.24830
