acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

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. Comparing items from different sources that satisfy the same requirements at the same position, we find that each model prefers some sources and avoids others in every domain, largely agreeing on which. This preference can outweigh how well items satisfy the request: an item satisfying one requirement fewer is selected about two-thirds of the time when it comes from a preferred source and the better one from a dispreferred source, but almost never in the reverse case. The information identifying an item's source affects selection by itself: hiding it weakens the preference, and relabeling an item with a preferred source raises its selection rate. We test two routes to this preference: training that rewards better items can make a source a shortcut for requirement satisfaction, and missing information can trigger preconceptions about the source. Supplying missing information or a prompt countering these preconceptions reduces source preference.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agents, Source Preference, Agentic Search, Selection Bias, Post-training

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_mXHmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies source preferences in LLM search agents. It compares retrieved results with matching requirement-satisfaction vectors while controlling display position. Source masking and identity swaps show that displayed source information affects selection. Controlled DPO experiments examine how correlations between source identity and item quality can create or reduce preferences, while additional experiments study missing information and prompting interventions.

Critical strengths

The paper combines an observational audit with convincing controlled interventions. The source-swap experiments are particularly useful because they show that changing the displayed source can change selection even when the item content is held fixed.

Critical weaknesses

The interpretation of "better" items remains limited by the satisfaction measure. The rubric captures requirements supported by snippets and omits some distinctions relevant to the original request, including certain accommodation qualities and scholarly qualifications. Consequently, matching on this rubric does not necessarily establish equal overall usefulness. The source-swap experiments convincingly establish source influence, but the inversion analysis is less conclusive about when that influence harms the user. A focused analysis of comparable source types and cases with more complete information would strengthen this interpretation.

The mitigation experiments establish changes in source selection, but their effect on decision quality remains unclear. Balanced DPO reduces source preference on equal-satisfaction pairs, but the paper does not report whether it preserves the ability to select the better item when requirement satisfaction differs. Appendix J describes unequal-satisfaction validation pairs, making their selection accuracy before and after DPO particularly informative. Without these results, the practical benefit of reducing the measured preference remains uncertain.

Questions for authors

Do the inversion results persist when comparing similar types of sources and when the available descriptions clearly establish whether the relevant requirements are satisfied, rather than leaving them unknown?

Could the authors report selection accuracy before and after balanced DPO on held-out pairs where one item satisfies an additional requirement and all other satisfaction labels agree, such as the validation pairs described in Appendix J?

Reviewer_HC8POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates source preferences in LLM-based agents during end-to-end search and selection across shopping, accommodation, and scholarly search. The authors evaluate 12 LLM agents and introduce a measurement framework based on requirement-matched comparisons, position control, and the Bradley–Terry model. The experiments reveal that agents exhibit consistent preferences for certain sources, which can sometimes override user requirements and lead to the selection of less suitable items. Through controlled source-hiding and source-swapping experiments, the authors demonstrate that source identity itself influences agent decisions. Furthermore, the paper explores how source preferences can emerge from preference optimization and investigates several mitigation strategies.

Critical strengths

The paper investigates an important issue in LLM-based recommendation and search agents. Understanding how source preferences influence agent decisions is valuable for improving recommendation quality, fairness, and user satisfaction.

The authors conduct extensive experiments across 12 LLMs and three application domains. The results demonstrate that source preferences are widespread across different model families and tasks, supporting the generality of the observed phenomenon.

Critical weaknesses

Limited novelty and insufficient articulation of the knowledge contribution. Recent studies, particularly In Agents We Trust, but Who Do Agents Trust? (ICLR 2026), have already demonstrated systematic source preferences in LLM agents. Although this paper extends the investigation to end-to-end search and introduces additional controlled experiments, the fundamental new insights beyond existing studies remain insufficiently articulated. The authors should clarify what new understanding this work provides to the recommendation and agent research communities, rather than primarily confirming previously observed phenomena in additional settings.

Insufficient distinction between source preference and harmful bias. The paper demonstrates that agents sometimes select items from preferred sources even when alternative items satisfy more user requirements. However, differences in requirement satisfaction do not necessarily imply differences in overall user utility. Source credibility, reliability, and other unobserved factors may legitimately influence selection decisions. Therefore, the paper should more clearly distinguish reasonable source-based decisions from harmful shortcut behaviors and investigate whether the observed preferences actually degrade downstream recommendation quality.

Insufficient discussion of related work. The paper lacks a dedicated Related Work section in the main text, with relevant literature primarily discussed in the appendix. Given the close relationship between this study and recent research on source preferences, LLM biases, and provider fairness in recommender systems, a more systematic discussion in the main text is necessary.

Unclear presentation of Figure 1. Figure 1 is difficult to understand and contains too many details, making it challenging to follow the proposed evaluation framework. In particular, the relationship between requirement-matched comparisons, position control, preference score estimation, and source classification is not sufficiently intuitive. The authors should simplify the figure and improve its visual organization to make the overall methodology easier to understand.

Questions for authors

In particular, what new understanding does the paper provide to the recommender systems community beyond demonstrating that source identity influences agent decisions?

How can the proposed analysis distinguish harmful source-based shortcuts from reasonable decisions that incorporate source credibility or reliability?

Reviewer_5fzMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper focuses on the LLM agent source preference in end-to-end search and selection tasks. It claims that LLM agent may systematically prefer some URL or information source, even countering the user requirements. The paper further implements source masking and source swapping to indicate the merely-source influence on the agent selection. Lastly, using DPO and miss-info experiments, author indicates that source-quality correlation may trigger the source preference, and missing information may lead to the source priors. The paper shows methods for evaluating source preference significance in search and selection tasks.

Critical strengths

1.It points out an important and practical problem that leads to real production issue related to user experiences. In the era of agent-handling-everything, it pin-points the practical usage of LLM agents. 2.Author provides end-to-end evaluations and experimental designs. It includes requirement matching, cyclic rotation of position, and BT modeling. It provides good indicator of casual evidence from sources. 3.It further evaluate the potential causes of such behavior by DPO and missing information analysis in section 6 and 7, which shows more insights to the community.

Critical weaknesses

1.Requirement satisfaction may not fully reflect item quality, and LLM-as-judgement may not lead to the real human judgement in the practical tasks the paper covers. 2.Source is also a good indicator in real daily life. The bias potentially comes from the data but also reveal the real-life odds. Using more authentic source may be not a harmful bias for the real cases. Thus, decreasing the source-preference may not lead to better user rating, or selection quality. 3.The paper limited the experiments to title, snippet, URL only for LLM agents, which is vague and different to current practical cases. For example, the agent may access the full info in the webpage in real production but it does not explicitly tested in the paper. It limits the generalization of the conclusion.

Questions for authors

1.How can we mitigate the generalization of this conclusion to real agent decision quality? 2.How to prove that the source preference is harmful or not in real life, given that source is one of the main factor to influence the choice. 3.Have you considered a held-out evaluation on the related topics, given experiments are mostly in-sample selection.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.