FPSRG turns personality profiles into a searchable catalog: five-level vectors plus prompts cover 98.7% of 25–75 targets
Synopsis
FPSRG replaces training one adapter per profile with shared first-person post-training plus a small set of retrievable directions. With the model frozen, it measures complete multidimensional scores for coefficient combinations and stores them in a catalog searched by target scores. In the five-dimensional non-thinking instance, the five-level vector catalog covers 92.2% of fixed targets in the 25–75 region, rising to 98.7% when combined with the prompt catalog.
Interpretation
Heterogeneous user simulation no longer needs a separately trained model for every profile; all profiles share one first-person post-trained model, and a small set of learned direction vectors supplies the attribute differences. Prompts can describe high or low tendencies, but degrees such as very, moderately and slightly still need score calibration, and training a LoRA for every profile brings repeated optimization and storage as combinations grow. With matched data, layers and norms, refinement raises the mean five-axis composite control span from 30.75 to 40.45, improving all five dimensions.
The model is first taught to judge and answer as the participant, then, with it frozen, complete score vectors are measured for direction-coefficient combinations and stored as a catalog retrievable by target scores. Coefficients and scores occupy different coordinate systems, so raising an extraversion coefficient may also change agreeableness; retrieval therefore uses the complete measured response rather than treating a coefficient as its trait score. The five-dimensional instance uses Qwen3.5-27B with a shared LoRA trained for one pass over 1,000 participant-perspective examples and then frozen; the five directions together give 3,125 configurations, and four-part filtering retains 1,854 nodes.
Catalog density and prompts expand the reachable region in complementary ways without retraining directions. At a fixed coefficient range, three-, five- and nine-level catalogs cover 71.2%, 92.2% and 94.6%, showing that denser measurement buys precision. Five-level vectors plus five-level prompts use 6,250 configurations and reach 98.7% broad coverage in the 25–75 region, above the 96.5% of 16,807 seven-level vectors alone.
Perspective
The results describe measured reachable regions under a fixed assessment protocol: the five- and ten-dimensional non-thinking instances use Qwen3.5-27B, the six-dimensional thinking instance uses Huihui Qwen3.8-27B-abliterated, and coverage and retrieval error are tied to each setting's items, encodings, sampling and injection. The Chinese PersonalityBench check only observes directional differences between high and low configurations in free responses and is not ranked against the original leaderboard.
The authors state that these scores are model response coordinates under specified tasks, not diagnoses of real individuals, and that generated reasoning need not faithfully describe hidden computation. Changed models, items or scoring rules require recalibration or independent checks, and thinking and non-thinking modes use separate catalogs.
