As LLM-powered agents increasingly make decisions on our behalf these days like making purchases, travel plans, medical choices, understanding how they choose matters as much as whether they succeed or not. This project introduced a framework for probing agentic choice through controlled manipulations of pricing, ratings, and psychological nudges in a realistic web-based shopping environment. The work found that agent decisions shift predictably and substantially in response to these cues, which showed that agents are really strongly biased choosers even without the cognitive constraints that produce human biases.
My contribution focused on exploring how model-level parameters (temperature, sampling strategies) in Google's Gemma family affect susceptibility to these nudges. We aimed to understand whether choice biases are artifacts of generation settings or deeper properties of the models themselves.