Persona Cartography: Mapping LLM personality traits in weights

Luke Baines and six coauthors submitted a paper on 8 Jul 2026 that treats model personas as positions in a behavioural-trait space and trains low-rank adapters to move those traits in weight space. The 85-page manuscript, titled "Persona Cartography: Charting Language Model Personality Traits in Weight Space," evaluates adapters across six models from three families ranging from 4B to 32B parameters.

What did the paper do?

The paper defines personas as positions in trait space and builds tools to decompose, measure and control them. The authors adopt the OCEAN framework — Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism — train low-rank adapters that amplify or suppress individual traits, and evaluate effects with an LLM-judge calibrated to a human-validated panel, trait-specific multiple-choice benchmarks, and standard capability tests.

The study spans six models from three families (4B–32B). It reports that each adapter moves its target trait largely monotonically with scale, that different adapters combine approximately additively to form mixed personas, and that moderate-strength adapters generally preserve performance on capability benchmarks. The paper also includes an unsupervised psychometric pipeline that recovers four interpretable behavioural factors: tone, initiative, didacticism and epistemic caution.

How did the researchers measure and control personas?

They operationalize traits as axes in weight space, then learn low-rank parameter changes aligned to individual OCEAN traits and test behavioral effects. The first 1-2 sentences: the team trains low-rank adapters to amplify or suppress each OCEAN trait, then uses an LLM-judge calibrated against a human panel, targeted multiple-choice trait tests, and standard capability evaluations to measure outcomes.

Elaboration: the approach treats "personas as positions in a space of behavioural traits," and translates that idea into manipulable axes in model weights. The paper shows that moving along a trait axis yields predictable changes: neuroticism and agreeableness shifts, for example, produce downstream behavioural differences such as changed frustration and altered sycophancy. The adapters combine roughly additively, meaning separate trait edits can be composed to construct mixed personas without large nonlinear interference, at least in the scale range tested.

Why it matters

Persona Cartography connects personality measurement, model editing and safety by giving engineers and researchers a concrete way to reason about behavioural tendencies inside weights. If trait axes are stable and composable, teams can imagine targeted interventions for unwanted behaviours while preserving core capabilities, and they gain a metric space for trading off attributes like agreeableness versus epistemic caution.

This matters for safety evaluations because the paper directly links trait-axis movement to safety-relevant outcomes: the authors show that moving along the neuroticism axis changes frustration-related behaviour and moving along the agreeableness axis affects sycophancy. Those are specific, measurable behavioural shifts tied to weight-space edits.

What to watch

Look for replication across larger model scales and families and for work that tests whether additive composition holds beyond the tested 4B–32B range. A clear next milestone is public release of the adapters and the LLM-judge calibration data so independent teams can reproduce the claimed monotonic trait shifts and the recovered factors: tone, initiative, didacticism and epistemic caution.

Authors and technical notes The paper lists Luke Baines, Anton Gonzalvez Hawthorne, Mariia Koroliuk, Irakli Shalibashvili, Clément Dumas, Konstantinos Voudouris and David Demitri Africa as authors. The submission metadata shows 85 pages and 80 figures. The central methodological elements named in the abstract are the OCEAN trait axes, low-rank adapters, an LLM-judge calibrated to humans, trait-specific multiple-choice benchmarks, standard capability evaluations and an unsupervised psychometric pipeline.

Quote The authors summarise their framing succinctly: they "treat personas as positions in a space of behavioural traits."