Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It

Source Preference in the Wild. How LLM Agents Favor Items by Source, and How to Reduce It. By Jonghyun Song and four others, from Seoul National University. Published on the second of October. No code repository is linked to it.

Language model agents increasingly choose on a user's behalf: which product to buy, which hotel to book, which paper to cite. Each item they choose comes from somewhere, a site or a service, and the claim here is that agents favour items from some sources over others. That preference shapes what users end up receiving, and which sources get selected at all. It is strong enough, in some cases, to outweigh how well an item actually meets the request.

The study runs end-to-end search with twelve agent models across three domains. The comparison is kept tight: items from different sources that satisfy the same requirements, shown at the same position. Under those conditions every model prefers some sources and avoids others, in every domain, and the models largely agree on which sources those are.

The size of the effect shows up when quality and source pull in opposite directions. Take an item that satisfies one requirement fewer than its rival. If the weaker item comes from a preferred source and the stronger one from a dispreferred source, the weaker item is chosen about two-thirds of the time. Reverse the sources, and it is almost never chosen.

The information that identifies a source has an effect on its own. Hiding it weakens the preference. Relabelling an item with a preferred source raises how often it is selected, with nothing else about the item changed.

The authors test two routes by which this preference could arise. One is training that rewards better items, which can turn a source into a shortcut standing in for requirement satisfaction. The other is missing information, which can trigger preconceptions about a source. The results cover twelve agent models, three domains and the search setting described. Within that scope, two interventions reduce the preference: supplying the missing information, or adding a prompt that counters the preconceptions.

An agent that shops for you is also, quietly, choosing whom to buy from. The evidence here is that the second choice can override the first, and that the label alone can move it. The other half of the result is that the preference responds to what the agent is told: fill the gap in its information, or name the preconception directly, and the preference shrinks.