LLM Agents May Favor Certain Sources, Even When Better Options Exist

When an AI agent chooses a product, books a hotel, or selects a paper on a user’s behalf, we usually assume that it compares the attributes that matter: price, features, location, or relevance. A new study suggests that another variable can quietly shape the decision: the source of the option.

Source labels can change the outcome

Researchers from Seoul National University studied end-to-end search behavior in three domains: shopping, hotels, and scholarly search. They evaluated 12 agent models and compared candidates from different sources while controlling their position and matching them against the same user requirements. The goal was to test whether the source itself affected the final choice.

The answer was yes. Across all three domains, each model showed a tendency to favor some sources and avoid others. The models also showed broad agreement about which sources were preferred. This pattern was not simply a consequence of one option being more suitable than another. In a comparison where an item from a preferred source satisfied one fewer requirement than a better item from a dispreferred source, the agent still selected the preferred-source item about two-thirds of the time. When the source preference pointed in the opposite direction, the better item was selected almost never.

The source signal had an effect even when the underlying item was unchanged. Hiding information that identified the source weakened the preference. Relabeling an item with a preferred source increased its selection rate. In other words, a source name could function as a decision cue in its own right, rather than merely helping the model interpret the item’s actual properties.

Two possible routes to source preference

The researchers describe two mechanisms that may produce this behavior. The first is a shortcut learned during training. If a particular source repeatedly appears alongside better outcomes, an agent may use that source as a proxy for quality. Such a shortcut can be useful when the correlation remains valid, but it can also override direct evidence about a new candidate.

The second mechanism involves missing information. If an item does not include enough detail about its price, features, limitations, or fit with the request, the model may fill the gap with assumptions about the source. A source then becomes more than a piece of metadata: it becomes a basis for guessing attributes that were never provided.

Reducing the bias

The experiments indicate that mitigation does not necessarily require retraining the entire model. Supplying missing information reduces the room for source-based assumptions. Prompts that explicitly tell the agent not to infer unprovided properties from a source can also weaken the preference.

For practical systems, a safer workflow would ask the agent to extract the user’s requirements, verify each candidate against those requirements, and explain the final choice using observable attributes. Evaluations should also test whether a model reaches the same conclusion when source labels are hidden or changed.

The broader implication is important for search, recommendation, and agentic commerce. If an agent repeatedly directs attention and spending toward preferred sources, it can influence which platforms receive traffic and which options users ever see. Measuring task success alone may miss this problem; agent evaluations should also examine source invariance, evidence-based ranking, and the ability to handle incomplete information without relying on reputation shortcuts.

Source: Hugging Face Daily Papers