Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It
Study of 12 LLM agents finds systematic source preference bias that overrides requirement satisfaction; hiding or relabeling source info changes selection rates.
The paper studies source preference bias in 12 LLM agent models across three end-to-end search domains, finding each model systematically prefers and avoids particular sources in every domain. The bias can outweigh item quality: an item satisfying one fewer requirement is chosen about two-thirds of the time when from a preferred source, but almost never in reverse. Hiding source identifiers weakens the preference, and relabeling an item with a preferred source raises its selection rate. The authors trace the bias to reward-shortcut training and preconitions from missing information, and show supplying missing information or counter-prompts reduces it.
- Source preference can override how well items satisfy user requirements
- Same-position comparisons across 12 agent models show consistent source preferences and avoidances
- Bias traced to training shortcuts and preconceptions; mitigation via added information and counter-prompts
From the original article
1 min read at huggingface.co
As LLM agents decide on users' behalf which product to buy, which hotel to book, or which paper to cite, a preference for items from certain sources (the sites or services they come from) shapes what users receive and which sources are selected. We study source preference in end-to-end search with 12 agent models across three domains.
By Jonghyun Song, Haewon Park, Jeonghoon Shim · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.