Controlling LLM Behavioral Styles and Profiles

Summary & Methodology Analysis

Large language models operate in interactive settings where behavioral styles impact user experience and safety. Existing personality studies rely on unstable first-person self-report questionnaires poorly grounded in concrete model behavior. To address this, the researchers constructed 3,200 contrastive behavioral scenarios spanning 20 behavioral patterns and four prompt registers, including first-person choice, daily advice, task advice, and task execution. These scenarios are grounded in validated psychometric facets such as BFI-2, DOSPERT, and HEXACO. Model responses are parsed into low or high pole behavioral modes using a response parser to estimate multidimensional behavioral profiles. Questionnaire self-reports and behavioral profiles exhibit an average gap of 22.7 percentage points across model subdomain pairs, while split-half analyses of behavioral profiles within registers show a mean split-half correlation of 0.933.

The technical approach builds activation space Behavioral Mode Axes, known as BMAs, by computing mean-difference vectors of hidden states from contrastive behavioral traces. The authors also extract thought derived BMAs from intermediate behavioral rationales and compare them with response derived BMAs extracted from final responses. Cross register profile correlations average 0.76 across the nine profile models, ranging from 0.37 to 0.97. Models evaluated include Llama-3.1-8B-Instruct, Llama-3.1-70B-Instruct, Qwen2.5-7B-Instruct, Qwen2.5-14B-Instruct, Qwen2.5-32B-Instruct, Gemma-2-2B-it, Gemma-2-9B-it, and Gemma-2-27B-it, along with datasets like LMSYS-Chat-1M and The Pile. To alter model behavior during execution, the system steers model behavior at inference time by adding a scaled BMA vector to activations within identified Behavioral Control Layer, or BCL, bands.

Despite these advancements, the paper notes key limitations. Response-derived BMAs are more prone to trait drift, which conflates behavioral patterns with output patterns and activates mechanisms unrelated to the intended style. Additionally, the same model's 20-dimensional behavioral profile only partially preserves its shape across different interaction registers. The paper does not specify precise hardware requirements, latency numbers, or dollar costs associated with applying these vector additions during inference.