analiza aiqbench
This study demonstrates that LLM self-reports on personality traits lack predictive validity for actual observed behavior, even across 25 models, revealing a significant gap between internal descriptions and external actions. The findings suggest that current self-report metrics may share variance with other LLM judges rather than reflecting human-observable traits, posing a risk for alignment evaluation pipelines. Researchers should treat LLM-generated personality profiles as diagnostic probes for internal consistency rather than reliable indicators of functional behavior.
An LLM-Native Psychometric Instrument Does Not Predict LLM Behavior: Evidence Across 25 Models
arXiv:2606.09843v1 Announce Type: cross Abstract: Large language models (LLMs) produce stable self-reports on personality inventories, but these self-reports do not predict observed behavior. Whether this gap reflects a mismatch between LLMs and human trait constructs, or a deeper property of LLM self-report itself, has been unresolved. We constructed the first psychometric instrument whose constructs are derived bottom-up from LLM behavioral affordances via exploratory factor analysis (EFA). We administered 300 items (240 direct Likert + 60 scenario-based) spanning 12 candidate behavioral dimensions to 25 LLMs across 17 model families, each item administered 30 times. EFA yielded a 5-factor structure -- Responsiveness, Deference, Boldness, Guardedness, and Verbosity -- with excellent split-half replicability (all Tucker $\phi \geq .957$) and internal consistency (all $\alpha \geq .930$). To test predictive validity, we collected 2,500 open-ended behavioral samples rated by 151 human raters and a three-judge LLM ensemble. Human and judge ratings agreed ($\bar{r} = .51$), but neither tracked self-report: self-report--human $\bar{r} = -.01$, self-report--judge $\bar{r} = .13$, with no factor-level self-report--human CI excluding zero. On Responsiveness, self-report correlated with LLM judges ($r = .53$) but not humans ($r = .04$), even though humans and judges agreed ($r = .59$) -- indicating self-report items and LLM judges share variance that human observers do not, a confound invisible to within-ensemble reliability checks. We release the instrument as a diagnostic probe for alignment-shaped self-description and a concrete risk factor for LLM-as-judge pipelines.
Abstract:Large language models (LLMs) produce stable self-reports on personality inventories, but these self-reports do not predict observed behavior. Whether this gap reflects a mismatch between LLMs and human trait constructs, or a deeper property of LLM self-report itself, has been unresolved. We constructed the first psychometric instrument whose constructs are derived bottom-up from LLM behavioral affordances via exploratory factor analysis (EFA). We administered 300 items (240 direct Likert + 60 scenario-based) spanning 12 candidate behavioral dimensions to 25 LLMs across 17 model families, each item administered 30 times. EFA yielded a 5-factor structure -- Responsiveness, Deference, Boldness, Guardedness, and Verbosity -- with excellent split-half replicability (all Tucker $\phi \geq .957$) and internal consistency (all $\alpha \geq .930$). To test predictive validity, we collected 2,500 open-ended behavioral samples rated by 151 human raters and a three-judge LLM ensemble. Human and judge ratings agreed ($\bar{r} = .51$), but neither tracked self-report: self-report--human $\bar{r} = -.01$, self-report--judge $\bar{r} = .13$, with no factor-level self-report--human CI excluding zero. On Responsiveness, self-report correlated with LLM judges ($r = .53$) but not humans ($r = .04$), even though humans and judges agreed ($r = .59$) -- indicating self-report items and LLM judges share variance that human observers do not, a confound invisible to within-ensemble reliability checks. We release the instrument as a diagnostic probe for alignment-shaped self-description and a concrete risk factor for LLM-as-judge pipelines.
Human-Computer Interaction (cs.HC); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2606.09843 [cs.HC]
(or arXiv:2606.09843v1 [cs.HC] for this version)
https://doi.org/10.48550/arXiv.2606.09843 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 328 cuvinte
Subiecte conexe
An LLM-Native Psychometric Instrument Reveals a Self-Report--Behavior Gap Across 25 Models
arXiv · 8 oct. 2026
Improving CoT Monitorability of Evaluation Awareness via Verbalization Training
LessWrong · 2 oct. 2026
Deconfusing Corrigibility
AI Alignment Forum · 17 oct. 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026