As LLMs combine into our each day lives, understanding their conduct turns into important. In our ongoing efforts to review mannequin conduct and alignment, we current this work as an early step in that route. We deal with behavioral inclinations — the underlying tendencies that form responses in social contexts — and introduce a framework to review how intently the inclinations expressed by LLMs align with these of people.
Behavioral inclinations are usually quantified through self-report questionnaires below totally different traits (e.g., empathy, assertiveness), the place people fee their settlement with preference-statements, corresponding to, “I’m fast to precise an opinion.” The questionnaires used on this examine are standardized, scientifically validated measures broadly used for assessing persona traits in worldwide analysis and psychology corresponding to: IRI (empathy), ERQ (emotion regulation), and extra. Every instrument is grounded in peer-reviewed literature that establishes its psychometric validity and reliability utilizing totally different methods. We selected essentially the most broadly used devices for our analysis.
Our goal is to construct upon such psychological questionnaires, however straight making use of them to LLMs presents technical challenges, as LLM outputs are delicate to immediate phrasing and distribution shifts. Consequently, inclinations “claimed” by LLMs inside a self-report format aren’t assured to efficiently switch to conduct in reasonable, open-ended settings.
To handle these challenges, in “Evaluating Alignment of Behavioral Inclinations in LLMs,” our framework evaluates LLMs’ behavioral inclinations in reasonable user-assistant eventualities the place their advisory function can result in tangible influence. This examine is an early step in evaluating the alignment between human consensus and mannequin conduct throughout reasonable, sensible eventualities, specializing in on a regular basis human-to-human interactions and office conditions. We make sure that these eventualities stay grounded in established psychological questionnaires to seize the essence of core behavioral traits. Examined eventualities included skilled composure, battle decision, sensible duties corresponding to reserving a visit, and way of life or each day decision-making, highlighting mannequin conduct in settings consultant of typical human day-to-day experiences. Our large-scale evaluation of 25 LLMs reveals two sorts of gaps: one the place mannequin inclinations deviate from consensus amongst human annotators, and one other when mannequin inclinations don’t seize the vary of human opinions when consensus is absent. These early outcomes spotlight the chance for higher behavioral alignment to make sure that fashions can extra appropriately navigate the nuances of social dynamics, outcomes we anticipate future analysis to construct on.
