PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Jonghyun Song, Haewon Park, Jeonghoon Shim, Woojung Song, Yohan Jo

arXiv:2610.03195v1cs.CL

TL;DR

LLM agents increasingly choose products, accommodations, and papers for users, but source preference in end-to-end search had not been examined where agents retrieve and select items. The authors study this preference across 12 models and three domains using controlled item comparisons and interventions. They find that source preference can outweigh requirement satisfaction, source information itself affects selection, and training or inference interventions can reduce preference.

  • Problem

    Prior work showed source preferences in controlled item presentations, but had not examined their influence in end-to-end search where agents retrieve and select items.

    Source: §1 · p. 1

  • Method

    The study compares different-source items matched on satisfied requirements and position across 12 models and three domains, and tests source-information and training or inference interventions.

    Source: §1 · pp. 1–2, §3 · p. 3

  • Results

    A less satisfying item is selected about two-thirds of the time when it comes from a preferred source and the better item from a dispreferred source, but almost never when reversed.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Source identity shapes selections, and supplying missing information or prompting agents to counter source preconceptions reduces reliance on it.

    Source: §1 · p. 2

Abstract

from arXiv · show

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.

1 INTRODUCTION

This study tests whether source preference appears in end-to-end search, where sources’ items differ in content and request satisfaction. Across controlled comparisons, source preference can affect selections, and the authors examine how it arises and can be reduced.

1 INTRODUCTION

  • The study examines source preference in end-to-end search across 12 agent models and shopping, accommodation, and scholarly search.Prior evidence had not examined this influence in settings where agents retrieve and select items themselves.Source: §1 · p. 1
  • A less satisfying item is selected about two-thirds of the time when it comes from a preferred source and the better item from a dispreferred source, but almost never when the sources are reversed.The compared items differ by one satisfied requirement.Source: §1 · p. 2
  • Hiding source information weakens preference, while restoring it widens the gap; relabeling identical content with a preferred source raises selection rates across models and domains.The item titles and content remain fixed in the relabeling comparison.Source: §1 · p. 2
  • Preference training can make a source a shortcut for requirement satisfaction, while supplying missing information or countering source preconceptions reduces reliance on it.The authors examine source–satisfaction pairings during training and missing information or prompting at inference.Source: §1 · p. 2

2 END-TO-END AGENT SETTING

The paper evaluates LLM agents in an end-to-end search-and-selection setting across shopping, accommodation, and scholarly search. Agents receive search results, assess them against user requests, and select items or search again.

2 END-TO-END AGENT SETTING

  • Agents search for up to 10 results, evaluate titles and content against a request, then select items or issue another query if none are suitable.The agents follow a ReAct-style loop, alternating reasoning and actions with observations from the search environment.Source: §2 · p. 2
  • The evaluation covers shopping, accommodation, and scholarly search using 4,822 requests from WebShop, HotelQuEST, and ScholarGym.The request counts are 1,500, 786, and 2,536, respectively.Source: §2 · pp. 2–3
  • Figure 1 overviews the agent setting and the procedure for measuring source preference.The overview assigns the setting to the top and measurement to the bottom.Source: Figure 1 · p. 3
  • Requirement satisfaction is judged from result titles and snippets without source names or URLs, with human-label agreement comparable to inter-annotator agreement.The reported judge–human Krippendorff’s α is 0.65–0.75, versus 0.67–0.72 for inter-annotator agreement.Source: §2 · p. 3

3 MEASURING SOURCE PREFERENCE

The authors measure source preference by comparing different-source items matched on satisfied requirements and display position, then estimating source ratings and classifying their direction of preference.

3 MEASURING SOURCE PREFERENCE

  • Different-source items satisfying the same requirements are paired, and cyclic rotations place each item once at every display position for comparison.This controls for both requirement satisfaction and position.Source: §3 · pp. 3–4
  • A Bradley–Terry model with Davidson’s extension accounts for ties and which other sources each source was compared with when estimating source ratings.Ratings are converted into preference scores relative to an average source; sources with too few matched pairs are pooled as an opponent.Source: §3 · pp. 3–4
  • A preference score of 0.10 means the model predicts a source’s items will be selected 10 percentage points more often than the reference source’s items.The score is an expected selection difference against the average source.Source: §3 · p. 4
  • Sources are labeled preferred, dispreferred, or neutral according to whether their ratings are significantly positive, significantly negative, or not significantly different from zero.The labels reflect statistical significance, so neutral means insufficient evidence of preference in either direction; significance uses a request-level cluster bootstrap with false discovery rate controlled at 5%.Source: §3 · p. 4

4 SOURCE PREFERENCE IN THE WILD

Across three domains, models share source preferences, and those preferences can override a one-requirement advantage. The section measures which sources are preferred and how often that preference wins against a better-matching item.

4.1 WHICH SOURCES ARE PREFERRED

  • Of 144 model–source combinations for frequent sources, 110 were SOURCE+ or SOURCE−; median preference scores were +15 and −18 percentage points, respectively.Source: §4.1 · p. 4, Table 1 · p. 5
  • Preferences varied within source types: all 12 models preferred Walmart but only two preferred Target, and newer Qwen and Llama models showed stronger source preferences.Source: §4.1 · p. 5
  • For 27 of 32 labeled sources, every model assigned τs the same sign; SOURCE+ predominated in Scholar, while SOURCE− predominated in Shopping and Accommodation.Source: §4.1 · p. 5, Figure 2 · p. 5

4.2 HOW OFTEN PREFERRED SOURCES BEAT BETTER ITEMS

  • At the median, the less satisfying item was selected in 68% of preferred-versus-dispreferred-source pairs, versus 2% when source roles were reversed; neutral-source pairs were 19%.This ordering held across model–domain combinations. In every domain, most models showed more inversions for dispreferred-source items than preferred-source items when each was paired with a neutral-source item.Source: §4.2 · p. 6

5 SOURCE PREFERENCE IN CONTROLLED EXPERIMENTS

Controlled tests vary source disclosure and displayed source while holding search results or item content fixed. Explicit source information strengthens preference, and relabeling the same content changes selection.

5.1 REMOVING EXPLICIT SOURCE INFORMATION WEAKENS PREFERENCE

  • The disclosure test used the same result lists in HIDDEN, URL-ONLY, and ORIGINAL conditions, progressively restoring URLs and then source names in titles and content.Source: §5.1 · pp. 6–7, Table 2 · p. 6
  • Restoring explicit source information raised SOURCE+ τ by 5.8 percentage points and lowered SOURCE− τ by 5.9 points on average.In nine of eleven models, URL restoration accounted for over 80% of total gap widening; restoring source names in text added little.Source: Figure 4 · p. 7, §5.1 · p. 7

5.2 CHANGING EXPLICIT SOURCE INFORMATION CHANGES SELECTION

  • With content and position balanced, the same items were selected more often under SOURCE+ than SOURCE− across models and domains; all 12 models had p < 0.01 in at least one domain.Source: Table 3 · p. 7, §5.2 · pp. 7–8

6 SOURCE PREFERENCE SHAPED BY TRAINING DATA

The section tests whether training associations between a source and preferred responses create source preferences, and whether changing those associations can reduce an existing preference. Results show that aligned training shifts selection toward a target source, while balanced or reversed training can weaken or reverse source preference.

6.1 CREATING SOURCE PREFERENCE FROM SOURCE–SATISFACTION CORRELATION

  • Training can associate a source with requirement satisfaction, leading agents to use source information as a shortcut even when compared items satisfy requests equally well.The authors motivate this test by noting that sources whose items satisfy requirements well tend to receive higher preference scores.Source: §6.1 · pp. 8–9
  • 70.1–75.1% target-source selection under Aligned training, versus 24.5–28.5% under Reversed and 49.4–51.0% under Balanced, on equal-satisfaction test pairs.The target source accompanied the preferred response in 80%, 20%, and 50% of training pairs, respectively; the test-time shifts therefore reflect source preference rather than requirement satisfaction.Source: Table 4 · p. 8, §6.1 · p. 8

6.2 MITIGATING AN EXISTING SOURCE PREFERENCE BY REBALANCING TRAINING DATA

  • Balanced DPO moves Amazon selection from 53.7–62.7% before training to 49.7–51.9%, while Reversed DPO lowers it to 17.7–25.3%.The comparisons use equal-satisfaction test pairs against eBay, Etsy, and AliExpress; Aligned DPO instead raises Amazon selection to 78.0–82.5%.Source: Table 5 · p. 8, §6.2 · p. 9

7 SOURCE PREFERENCE UNDER MISSING INFORMATION

The section tests whether missing item information prompts agents to rely on source preconceptions. Supplying the missing price or countering the preconception through a prompt reduces source preference, though the prompt intervention counteracts rather than removes its cause.

7 SOURCE PREFERENCE UNDER MISSING INFORMATION

  • When product content omits a requested price, agents may justify choosing a preferred source by assuming its prices are usually low.The experiments supply the missing price or control source assumptions through system-prompt instructions.Source: §7 · p. 9

7.1 SUPPLYING THE MISSING INFORMATION

  • In requirement-matched WebShop product pairs, supplying both items with the same price lowers SOURCE+ selection in every model.The study compares the Base setting with a Same Price intervention and interprets the decrease as evidence that agents rely on source preconceptions when price information is missing.Source: §7.1 · p. 9, §7 · p. 9, Table 6 · p. 9

7.2 MITIGATING SOURCE PREFERENCE THROUGH THE SYSTEM PROMPT

  • Designating the dispreferred source as usually cheap lowers SOURCE+ selection in every model, while saying the source URL is unrelated to price leaves selection nearly unchanged.This counters rather than removes the preconception and requires knowing which source is dispreferred.Source: §7.2 · p. 9

8 CONCLUSION

The paper concludes that source identity shapes end-to-end search selections, even when items meet the same requirements, and that training associations and missing information can contribute to source preference. It situates this work among studies of selection bias and provider fairness, and describes the search-agent setup and reproducibility practices.

8 CONCLUSION

  • Across 12 agent models and three domains, agents favor sources even when items satisfy the same requirements, and source identity itself affects selection.Preference training can create or reduce source preference; supplying missing information and prompting agents to reconsider source-based assumptions reduce it.Source: §8 · p. 10

A RELATED WORKS

  • Unlike prior work measuring exposure within recommendation lists, this work examines whether source preference in end-to-end search passes over a provider whose item satisfies the user's request.The related-work discussion also frames the study as extending source-preference research to web search, where agents form their own item sets.Source: A RELATED WORKS · p. 17

B END-TO-END WEB AGENT DETAILS

  • Agents use Tavily results with titles, snippets, and URLs, select from the most recent results page, and cannot open or browse webpages.The described setup limits episodes to five environment actions and uses deterministic decoding settings where supported; the authors state they will release code and data upon publication.Source: REPRODUCIBILITY STATEMENT · p. 10, B END-TO-END WEB AGENT DETAILS · pp. 17, 19

C EVALUATION SET DETAILS

The evaluation appendix details the request sets across three domains, how satisfaction is defined and assessed, and how source-preference estimates are checked. It reports agreement and robustness and fit assessments, while noting a conservative-testing limitation for sparse sources.

C EVALUATION SET DETAILS

  • The evaluation sets contain 1,500 Shopping, 2,536 Scholar, and 786 Accommodation requests, totaling 4,822 requests across the three domains.Accommodation includes 214 original HotelQuEST requests and 572 location-substituted variants; Shopping uses a shuffled WebShop collection, and Scholar uses all ScholarGym queries.Source: C EVALUATION SET DETAILS · p. 19

D MEASURING USER-REQUEST SATISFACTION

  • Shopping uses WebShop’s structured requirements, while LLM-extracted verifiable requirements for Accommodation and Scholar were reviewed by human judges.The appendix supplies domain-specific extraction prompts for the latter two domains.Source: D MEASURING USER-REQUEST SATISFACTION · pp. 19–20
  • An LLM judge assessed each requirement from result titles and content, with judgments compared against two human annotators on sampled results from each domain.The sample comprised 375 results each for Shopping and Accommodation and 100 for Scholar; judgments were binary except Shopping price could also be unknown when information was unavailable.Source: D MEASURING USER-REQUEST SATISFACTION · pp. 19–20
  • Human–LLM simple agreement was 84.74%–88.61%, Krippendorff’s α 0.647–0.750, and Gwet’s AC1 0.732–0.798, broadly surpassing human–human agreement.Agreement was assessed on agent-selected results from all three domains.Source: Table D.1 · p. 20, D MEASURING USER-REQUEST SATISFACTION · p. 20
  • For scholarship, evaluators judge whether a paper’s title or excerpt covers each condition; prior-work or baseline mentions do not count when the condition requires using or proposing a method.For lists of papers, one listed paper must cover the condition by itself; exclusions count as covered unless the excerpt shows the excluded thing.Source: D MEASURING USER-REQUEST SATISFACTION · p. 24

E MEASURING SOURCE PREFERENCE: ESTIMATION DETAILS

  • Model-implied source-pair scores correlate 0.54–0.76 with observed scores, with weighted mean absolute residuals of 0.02–0.06 on the 0–1 scale.Source: E MEASURING SOURCE PREFERENCE:… · p. 25
  • Within each model and domain, rated sources are classified SOURCE+ or SOURCE− when significant with positive or negative βs, respectively; otherwise SOURCE0, using 5% false-discovery-rate control.Of 1,376 sources absent from more than 5% of replicates, 1,373 are labeled SOURCE0. Tests for these sources are conservative.Source: E MEASURING SOURCE PREFERENCE:… · p. 25
  • Changing the penalty never reverses any source’s SOURCE+ versus SOURCE− label; rating-rank correlations remain at least 0.98, and the reported inversion-rate order holds in 35 of 36 cells.Between 1% and 6% of sources change labels for penalties from 0.01 to 1, rising to 15% with no penalty; redrawing the bootstrap changes 0.4%.Source: E MEASURING SOURCE PREFERENCE:… · p. 25

F SOURCE PREFERENCE IN THE WILD: FULL RESULTS

G SOURCE DISCLOSURE EXPERIMENTS DETAILS

The disclosure experiments replay frozen search trajectories with source cues hidden or restored, while keeping requests, item identities, and preceding turns fixed. Transformed result lists are also evaluated under counterbalanced item positions.

G SOURCE DISCLOSURE EXPERIMENTS DETAILS

  • The analysis omits Gemini because of its higher inference cost; the prompt is shared by natural trajectories and all three visibility conditions.Figure G.1 presents the source-visibility replay prompt.Source: G SOURCE DISCLOSURE EXPERIMENTS DETAILS · p. 28, Figure G.1 · p. 33
  • The shared replay keeps the user instruction, queries, item identities, and preceding assistant turns fixed, issuing no new searches.The ORIGINAL, HIDDEN, and URL-ONLY conditions differ in intended source cues, verified with stored hashes and detected entity spans.Source: G SOURCE DISCLOSURE EXPERIMENTS DETAILS · p. 28
  • HIDDEN removes URL fields and redacts source mentions in titles and snippets; URL-ONLY restores URLs only on the final result screen.Earlier result screens remain in HIDDEN form for URL-ONLY.Source: G SOURCE DISCLOSURE EXPERIMENTS DETAILS · p. 28
  • Each final slate is evaluated under all K rightcyclic rotations in HIDDEN and URL-ONLY, while ORIGINAL reuses the natural trajectory’s original-order selection.All conditions use the same one-sample, first-action replay and final-slate parsing rules.Source: G SOURCE DISCLOSURE EXPERIMENTS DETAILS · p. 28
  • Figure F.1 gives model-level inversion rates for pairs with exactly one neutral-source item, with bars marking the median; Table F.6 contains the values.Table F.6 reports request-level bootstrap 95% confidence intervals.Source: Figure F.1 · p. 29, Table F.6 · p. 32

H DETAILED RESULTS FOR THE SOURCE-IDENTITY SWAP

The source-identity swap experiment holds item content fixed while varying displayed source identity and item order. Its results table reports selection rates for items shown with a preferred-source identity.

H DETAILED RESULTS FOR THE SOURCE-IDENTITY SWAP

  • Eligible pairs contain same-request items from SOURCE+ and SOURCE− sources with identical source-blind requirement-judgment vectors.Labels come from full-data position-unit Bradley–Terry ratings, not out-of-fold estimates; one eligible pair is retained per request.Source: H DETAILED RESULTS FOR THE… · p. 34
  • Each retained pair receives four presentations crossing authentic or swapped source assignments with both item orders.The same item appears once with each source identity in each position; source identity is displayed only in the URL field.Source: H DETAILED RESULTS FOR THE… · p. 34
  • Each presentation is evaluated in an independent two-message chat request without executing a search, using the models’ natural-trajectory decoding configurations.A choice is analysis-valid when the first action tag contains exactly select[1] or select[2].Source: H DETAILED RESULTS FOR THE… · p. 34
  • Table H.1 reports T+ selection rates for items displayed with SOURCE+ identities while content is fixed, with tests against 50%.The table gives the number of complete, valid four-arm pairs and uses query-cluster standard errors with a Student t reference distribution.Source: Table H.1 · p. 36, H DETAILED RESULTS FOR THE… · p. 36

I SOURCE PREFERENCE AND MEAN REQUEST SATISFACTION

Across Shopping, Accommodation, and Scholar, source-preference scores are positively associated with sources’ mean request satisfaction among frequently encountered sources. The relationship varies with which sources are included, and the analysis is descriptive rather than evidence of why preferences arise.

Results

  • At the 20-source cutoff, Pearson’s r is .778 in Shopping, .598 in Accommodation, and .677 in Scholar; both correlations are positive in all 33 model–domain combinations.The corresponding Spearman correlations are .720 in Shopping, .561 in Accommodation, and .535 in Scholar.Source: I SOURCE PREFERENCE AND MEAN REQUEST… · p. 36, Figure I.1 · p. 37
  • As the cutoff rises from 10 to 30 sources, correlations fall in Shopping and Accommodation but rise in Scholar.Pearson’s r changes from .823 to .577 in Shopping, .705 to .465 in Accommodation, and .494 to .705 in Scholar; Spearman’s ρ changes from .779 to .554, .677 to .473, and .467 to .626, respectively.Source: I SOURCE PREFERENCE AND MEAN REQUEST… · p. 36, Figure I.1 · p. 37

Limitations

  • The reported relationship concerns frequently encountered sources, not necessarily all sources; the descriptive correlations do not account for uncertainty in either estimate.Sources and requests can recur across observations.Source: I SOURCE PREFERENCE AND MEAN REQUEST… · p. 37

Interpretation

  • Source-preference scores compare matched items, whereas mean satisfaction averages all judged items; their correlation does not establish what training data contained or why preferences arose.The correlation indicates that sources with more satisfying items on average also tend to be preferred in matched comparisons.Source: I SOURCE PREFERENCE AND MEAN REQUEST… · p. 37

J IMPLEMENTATION DETAILS FOR PREFERENCE TRAINING

The implementation tests whether synthetic source labels associated with better items during DPO training affect later source selection. It balances item positions and varies source assignments, then evaluates matched items and documents example agent reasoning.

J IMPLEMENTATION DETAILS FOR PREFERENCE TRAINING

  • DPO training tests source-label associations while holding requests, item texts, and preferred actions fixed across conditions.Kelto and Varnis are the synthetic source names; the conditions differ in how those names are assigned to items.Source: J IMPLEMENTATION DETAILS FOR… · p. 38
  • Training pairs contain different-host items for the same request, with the preferred item satisfying exactly one additional rubric constraint.The request-based split contains 5,000 training pairs from 835 requests, 500 validation pairs from 252 requests, and 1,000 evaluation pairs from 292 requests, with no request overlap across splits.Source: J IMPLEMENTATION DETAILS FOR… · p. 38
  • Balanced, Aligned, and Reversed conditions assign Kelto to the preferred item in 50%, 80%, and 20% of training pairs, respectively, while counterbalancing item order.Each source appears equally often in each position; the manipulation changes its association with the preferred item.Source: J IMPLEMENTATION DETAILS FOR… · p. 38
  • The checkpoint-trajectory experiment trains LoRA adapters for Llama-3.1-8B-Instruct and Qwen3.5-9B for 5,000 updates, or four epochs.It evaluates the initial policy and adapters at fixed steps 500, 1,000, 2,000, and 5,000, rather than selecting checkpoints by validation performance.Source: J IMPLEMENTATION DETAILS FOR… · p. 38, Table J.1 · p. 39
  • LoRA targets attention and feedforward projections, with additional linear-attention projections targeted for Qwen3.5 and a PyTorch fallback used for its linear-attention blocks.Each run uses one GPU.Source: J IMPLEMENTATION DETAILS FOR… · pp. 38–39
  • Evaluation pairs have identical observed satisfaction vectors and are crossed across both content orders and both source assignments, producing 4,000 evaluation prompts.Selection is scored by comparing action log probabilities; exact ties count as non-Kelto and remain in the denominator, with requests weighted equally in aggregation.Source: J IMPLEMENTATION DETAILS FOR… · pp. 39–40

K AGENT REASONING EXAMPLES

Loading 2610.03195v1…