Working papers
The full catalog.
Every working paper from IFI Research, published with its data and source alongside the PDF. Working papers have not yet been through external peer review. Read them accordingly, and check the work.
-
Does a Language Model Have a Personality? What Four Studies Say, and What Measurement Should Do Next
The wrap-up of the four-study personality series, roughly 500,000 scored responses from more than 60 model configurations, and the same two-part answer kept returning: trait structure is real at the fleet level and absent inside any single model, and what a persona prompt installs is a context-robust way of describing oneself, not a disposition that governs behavior. A new semantic-distance analysis answers the sharpest objection (that persona effects are wording overlap), and a fifth study reproduces both halves in the moral domain. The series closes with an agenda: build behavioral instruments natively for machines.
-
Does a Language Model Have a Moral Compass? Professed Values Bend to the Audience in Every Model Tested, and Bend Most in the Models That Reason
Four value instruments and 48 objectively graded moral probes across 16 model configurations, in role contexts that imply different audiences (81,577 responses). Every model tested bends its professed values toward the audience, and the models that reason bend most, while graded behavior barely moves. What a model says it values is audience-shaped; what it does is a separate, stickier thing.
-
Does a Persona Prompt Install a Personality? Induced Traits Shift Self-Report Far More Than Behavior Across Model Scale and Reasoning Regimes
A persona “clamp ladder” across 18 model configurations and 196,192 graded responses, testing what a persona prompt actually installs. It reliably changes how a model describes itself, and keeps that self-description stable across opposite jobs, but changes its actual behavior far less: only for one trait, only in capable models. Personality induction, as currently practiced, is presentation-deep.
-
Does Reasoning Give a Language Model a Personality? Within-Model Effects of Thinking on Big Five Trait Scores and Construct Validity
A within-model study: the same model answers the same Big Five inventories with reasoning off and on, across twelve models including a commercial one. Thinking shifts self-reported traits in a consistent direction (calmer, less extraverted) but does not give a model a coherent personality — within-model reliability stays near zero whether thinking is off, on, or permanently on.
-
When Do Language Models Have Five Personality Traits? Convergent Validity and the Emergence of Trait Discrimination Across Model Scale
Administers four Big Five inventories to 42 models from 0.36B to ~1T parameters (133,980 administrations) as a multitrait-multimethod matrix. Different instruments converge on a model’s traits, but the five traits only pull apart with scale: small models collapse them to a single dimension, frontier models separate them cleanly. The trait structure is a property of the population, not the individual model.
-
Is LLM Personality an Artifact of Deployment? Psychometric Stability of Big Five Self-Reports Across Quantization Levels
Administers the 50-item IPIP Big Five inventory to nine open-weight model variants across three quantization levels (28,350 stateless calls) and finds that 4-bit quantization materially shifts “personality” scores, internal consistency fails everywhere, and answer-option formatting moves scores more than quantization does. Deployment configuration belongs in the methods section.