What's this about?
LLMs are increasingly deployed as decision-making agents in settings requiring sophisticated exploration of their environment. This paper shows their in-context exploration is biased by semantic priors — what the options are *called* shapes which ones get tried, independent of how they actually perform.
What exactly is a "bandit" problem?
The name comes from slot machines — one-armed bandits. You face several levers with unknown payoffs and must decide, on every pull, whether to exploit the one that has done best so far or explore another that might be better. It's the cleanest formalization of the explore-exploit tradeoff, and it underlies recommendation, A/B testing, and clinical trial design. The classic assumption is that the arms are anonymous — you learn only from outcomes. The finding here is that LLM agents violate that: an arm labelled "premium option" gets explored differently from one labelled "option C," even when the payoffs are identical.
Why is this worth noticing?
- It's a concrete, measurable bias in a setting with clean theory, which makes it far more actionable than general claims about model prejudice.
- It has immediate practical consequences: an agent evaluating vendors, tools or strategies will explore unevenly based on names — meaning whoever writes the labels influences the outcome.
- It's not purely a defect. Semantic priors are useful when names carry real information; the problem is that the agent can't tell when they don't.