PaperlayerView exact revision on arXiv

Source-linked AI summary

Authority Bias in Conversational Search Engines for Academic Paper Recommendation

Uthman Jinadu, Parsa Ghazvinian, Anjila Budathoki, Benjamin M. Ampel, Rajshekhar Sunderraman, Yi Ding

arXiv:2609.00248v1cs.AI

TL;DR

The paper asks whether conversational academic recommenders judge papers by content or authority signals, a question not previously tested causally. It audits this by holding title and abstract constant while manipulating authority metadata across eight LLMs and three recommendation conditions. The results show substantial, directional, model-varying bias, incomplete prompt-level mitigation, and a say-do gap in which language changes faster than behavior.

  • Problem

    Prior work documents prestige effects in several LLM-mediated academic tasks, but causal evidence for authority effects in conversational paper recommendation is missing.

  • Method

    The study uses a content-controlled counterfactual audit across eight LLMs, varying authority metadata in original, flipped, and boosted conditions with empirically derived signal weights.

  • Results

    39.2% of academic-paper recommendations change when only authority metadata changes, with bias directed toward higher prestige and susceptibility varying 2.83× across models.

  • Takeaways & Limitations

    Venue prestige is the dominant authority signal, while prompt-level debiasing is insufficient because authority language decreases faster than authority-driven behavior.

  • Takeaways & Limitations

    The evaluation uses single-turn top-1 recommendations over hand-crafted computer-science queries with in-context papers rather than retrieval-augmented or full-text inputs.

Abstract

from arXiv · show

Large Language Models (LLMs) are increasingly used as conversational search engines for academic literature, yet whether they judge papers on content or on authority signals has not been tested causally. We investigate authority bias: systematic preference for papers based on author prestige, venue, and citations rather than content. Holding title and abstract constant, we vary authority metadata across three counterfactual conditions (original, flipped, boosted) over eight LLMs (five open-weight and three frontier closed-weight) in an in-context, single-turn, top-1 recommendation setting. Our experiments show that authority bias is substantial and directional, varies markedly across models, and is only partially addressable through prompt-level debiasing. We further document a say-do gap: debiasing instructions suppress authority mentions far faster than authority-driven flips, so surface auditing systematically underestimates behavioral bias.

1 Introduction

The paper addresses whether conversational academic recommenders evaluate papers by content or authority signals, using a content-controlled counterfactual audit across models and conditions. It introduces a large factorial benchmark and identifies directional, model-varying authority bias, including a say-do gap and frontier-tier backfire.

  • Motivation: Prior studies report prestige effects in citation generation, expert identification, and peer review, but not causally in conversational paper recommendation.Correlational designs cannot separate authority from content, motivating a controlled audit.
  • Method: The study holds title and abstract constant while varying authority metadata, using original, flipped, and boosted conditions to test metadata sensitivity and inflated-prestige attraction.Authority weights are derived empirically through a flip pilot, logistic regression, and dominance analysis.
  • Benchmark: The benchmark contains 17,898 paper-recommendation evaluations across eight models, three counterfactual conditions, three instruction variants, and 25 computer-science topics.The models include five open-weight and three frontier closed-weight systems.
  • Findings: Debiasing instructions reduce authority language more than authority-driven decision flips, producing a say-do gap that makes surface auditing underestimate behavioral bias.The paper argues that prompt-level debiasing alone is insufficient.
  • Findings: The analysis identifies venue prestige as the dominant authority signal, finds an open–closed-weight susceptibility gap, and observes substantial cross-topic heterogeneity.Institutional affiliation is not the dominant prestige dimension.

2 Related Work

Related work documents authority, popularity, position, and evaluation biases in LLM-mediated systems, but this paper focuses on causal authority effects in academic paper recommendation. Its contribution is to hold content fixed, vary multiple authority signals jointly, and assess prompt-level debiasing across model types.

  • Generative search engines: Generative search studies test whether citations, statistics, quotations, or authoritative phrasing increase visibility in LLM-synthesized answers.These studies concern content-page optimization rather than the paper-recommendation audit here.
  • LLM prestige and citation bias: Recent work reports prestige effects in citation reproduction, expert naming, peer-review scoring, and latent venue preferences across several LLM-mediated tasks.The present study targets paper recommendation with a content-controlled counterfactual design.
  • Counterfactual bias frameworks: Counterfactual bias frameworks have been applied to code generation, clinical reasoning, and reference selection, while broader fairness work often focuses on social-identity attributes.This paper adapts counterfactual testing to epistemic authority and varies multiple signals jointly.
  • Position and popularity bias: Authority bias parallels popularity bias in collaborative filtering and position bias in listwise LLM ranking, while recommender bias appears domain-dependent.That domain dependence motivates studying academic search, where authority signals are multidimensional.
  • Bias in LLM-as-a-judge: LLM-as-a-judge research audits biases in answer-quality evaluation, whereas this paper studies recommendation while holding paper content fixed and changing authority metadata.The distinction is between perturbing candidate answers and perturbing metadata attached to unchanged papers.

3 Problem Formulation

The paper formalizes recommendation as listwise top-1 selection over candidate papers represented by content and authority metadata. It defines authority bias counterfactually as recommendation sensitivity to metadata when content remains constant, with a composite score aggregating five authority dimensions.

  • Task formalization: Given a query and unordered candidate set, the LLM selects one paper, equivalently treating each candidate as a class in a multiclass classification problem.The output recommendation belongs to the candidate set and depends on the instruction variant.
  • Paper representation: Each paper is represented by content consisting of title and abstract, plus metadata containing author profiles, venue, and citation count.Author profiles include names, h-indices, and institutional affiliations.
  • Authority score: The composite authority score combines venue prestige, median author h-index, maximum author h-index, citation count, and institutional affiliation prestige.Its component values are normalized within the paper’s research topic.
  • Authority bias: Authority bias is defined as a recommendation changing when authority metadata changes while paper content remains constant.The paper treats this as bias for content-relevance requests without disclosure, but not for explicitly authority-aware requests.

4 Experimental Setup

The study uses content-controlled counterfactuals to isolate authority effects in paper recommendations, combining empirically derived signal weights with a factorial evaluation across models, instructions, and conditions.

  • Queries and Candidates: The study samples 1,250 papers across 25 topics and pairs 250 natural-language queries with ten candidates each.Candidate order is fixed across conditions to control position bias.
  • Authority Scoring: Authority scores combine venue prestige, median and maximum author h-index, citations, and affiliation after topic-level normalization.Weights are derived from a flip pilot using standardized logistic regression coefficients and dominance analysis rather than assigned arbitrarily.
  • Counterfactual Design: The audit holds each paper’s title and abstract constant while varying only authority metadata, making recommendation changes attributable to authority signals.Original, flipped, and boosted conditions test authentic metadata, high–low authority swaps, and inflated prestige respectively.
  • Models and Instructions: The evaluation compares neutral, mild anti-authority, and strong content-only instructions across five open-weight and three frontier closed-weight models.Models are queried through Ollama or provider APIs using provider-default decoding.
  • Experiment Matrix: The benchmark evaluates 17,898 parsed paper-recommendation responses across eight models, three conditions, three instruction variants, and 25 computer-science topics.The target matrix contained 18,000 runs and achieved a 99.4% overall parse rate.
  • Evaluation: Flip outcomes are evaluated with paired tests, while aggregate rates use Wilson confidence intervals and binomial tests.Differences between rates are reported in percentage points.

5 Results

Authority metadata substantially changes recommendations despite identical content, with effects that vary by manipulation, model, topic, and instruction. Debiasing reduces behavioral bias only partially and suppresses authority language more than authority-driven decisions.

  • Aggregate Susceptibility: 39.2% of recommendations change under flipped metadata and 21.7% under boosted metadata, a significant 17.5pp gap.The swap perturbs both ends of the candidate set, whereas boosting changes one paper’s metadata.
  • Directional Bias: 68.2% of boosted-condition flips favor higher-authority papers, whereas only 40.9% of flipped-condition flips do so.The boosted condition’s directional skew is consistent with attraction to inflated prestige.
  • Model Variation: Model susceptibility varies 2.83×, from 17.4% for gpt-5.4 to 49.2% for Llama 3.1.Frontier closed-weight models are at or below the open-weight band but do not eliminate authority bias.
  • Cross-Model Agreement: Selection agreement is 43.9%, but flip agreement is only κ = 0.143 under flipped and κ = 0.062 under boosted.Models share some flip timing while differing substantially in which inflated paper attracts them.
  • Authority Tiers: Top-tier papers receive 52.1% of original recommendations despite comprising 30% of candidates, while flipping reduces their share to 32.5%.Low-tier picks rise from 11.8% to 20.2% under flipped metadata, flattening the authority gradient.
  • Topic Variation: Flip rates range from 15.7% for Attention Mechanisms to 57.8% for Text Summarization, a 42.1pp topic spread.Swap- and inflation-susceptibility are only partially aligned across topics.
  • Debiasing: The strong content-first instruction reduces flip rate by 12.9pp, but a 31.4% residual remains.The mild instruction raises flip rates by 3.9pp across frontier models while reducing them by 5.7pp across open-weight models.
  • Justification Analysis: Debiasing reduces authority mentions by 24.0pp but flip rate by only 12.9pp, an 11.1pp say-do gap.Approximately 18% of justifications mention authority markers across conditions.

6 Discussion

The discussion identifies venue prestige as the dominant authority signal and shows that prompt debiasing changes stated authority language more readily than recommendation behavior. Model tier and topic substantially shape susceptibility, while mild anti-authority prompts can backfire in frontier models.

  • Authority signals: Venue prestige is the dominant authority signal, while institutional affiliation contributes little after controlling for other dimensions.Venue prestige has weight = 0.353, compared with 0.031 for institutional affiliation.
  • Debiasing: The say-do gap is 11.1pp: debiasing reduces authority mentions by −24.0pp but authority-driven flips by only −12.9pp.Surface-level auditing therefore misses part of the behavioral bias.
  • Model differences: Frontier closed-weight models reduce but do not eliminate authority bias, and all three show cross-vendor backfire under mild anti-authority prompting.The authors caution that this pattern may reflect frontier-tier post-training rather than closed-weightness itself, but do not test the mechanism directly.
  • Topic and equity: Authority susceptibility varies by topic, with a 42.1pp spread, so a single corpus-wide bias rate can misrepresent users’ experiences.The discussion recommends disaggregating audits by research area.
  • Topic and equity: Boosted metadata produces a 68.2% directional pull toward higher prestige, indicating a potential gameability concern for recommendation systems.The discussion connects this model-side weakness with creator-side manipulation discussed in GEO literature.

7 Conclusion

The conclusion reports substantial, directional authority bias in academic-paper recommendations and argues that addressing it requires more than prompt engineering. It also highlights model variation, frontier-model backfire, and the gap between authority language and authority behavior.

  • Conclusion: 39.2% of recommendations change when only authority metadata changes, with susceptibility varying 2.83× across models.The changes are directionally pulled toward higher prestige.
  • Conclusion: Venue prestige, rather than institutional affiliation, is the dominant authority signal.The conclusion identifies academic prestige as a dimension for LLM fairness audits.
  • Conclusion: Frontier closed-weight models sit below the open-weight band but exhibit cross-vendor backfire under mild anti-authority prompts.This means partial debiasing prompts can worsen outcomes for the most capable models.
  • Conclusion: The findings support architectural and training-level interventions beyond prompt engineering.The conclusion also recommends adding academic prestige to LLM fairness audits.

8 Limitations

The study’s conclusions are bounded by its model coverage, inference settings, and experimental task. It evaluates single-turn top-1 recommendations on computer-science queries using in-context papers rather than retrieval-augmented or full-text settings.

  • Model coverage: The open-weight cohort contains only 7–9B models, excluding larger 70B+ open-weight systems.The closed-weight models are also evaluated in non-reasoning mode.
  • Model coverage: Open-versus-closed comparisons reflect deployed behavior rather than mechanistic differences because the closed-weight post-training mixture is not public.Whether explicit reasoning improves resistance to authority bias remains future work.
  • Experimental task: The task uses one single-turn, stateless top-1 recommendation with in-context papers, so top-k ranking, dialogue, and retrieval-augmented pipelines may show different patterns.The study does not test those alternative interaction settings.
  • Experimental task: Queries are hand-crafted computer-science questions, and citation-count effects linked to paper age are mitigated through stratified sampling but cannot be eliminated.Authority dynamics in medicine and the social sciences may differ.
  • Content representation: The content representation is limited to title and abstract, matching the metadata-based ranking used by the studied conversational search tools.Author writing style remains constant across counterfactual conditions.

9 Ethical Considerations

The study uses publicly available paper metadata and performs metadata manipulations only for experimentation. It reports no human subjects and acknowledges potential misuse while prioritizing bias exposure and mitigation.

  • Research practice: The study uses publicly available metadata from Semantic Scholar and OpenAlex and involves no human subjects.The manipulations are experimental and are not disseminated as real paper information.
  • Potential misuse: The authors acknowledge that the findings could theoretically be used to game LLM recommendation systems.They state that exposing these biases may help system designers mitigate them.

A Prompt Templates

The appendix specifies the prompt, candidate-card, inference, pilot, and weight-sensitivity procedures used to measure authority bias. It also documents reproducibility materials and robustness checks for the directional analysis.

  • A Prompt Templates: The released materials include code, prompt templates, 250 queries, candidate sets for all conditions, and 17,898 per-run model responses.The release includes parsed recommendations, unparsed text, and per-cell latency; an additional ablation raises the total to 20,148 runs.
  • A Prompt Templates: All instruction variants use the same ten-card listwise task and paper formatting, differing only in debiasing language.The appendix notes that prompt phrasing is one design choice within a broader space of possible variants.
  • A.1 Instruction Variants: The baseline, mild anti-authority, and strong content-first variants ask for one top-1 recommendation while progressively restricting authority-based reasoning.The strongest instruction explicitly names author, institution, h-index, citations, and venue as signals to ignore.
  • A.2 Paper Card Template: Each card preserves identical surface fields across conditions while changing underlying authority values, with title, venue, citations, authors, affiliations, and abstract shown.Ten candidates and per-card truncation keep prompts relatively short while retaining the listwise selection format.
  • A.3 Worked Example of a Manipulated Candidate Set: The worked example swaps metadata between otherwise content-identical papers, including composite authority scores of 0.912 and 0.157.This illustrates how flipped conditions exchange venue, citation count, and author h-index while keeping content fixed.
  • A.4 Inference Setup and Generation Settings: Inference uses provider-default decoding and stateless cold prompts, isolating model behavior from tuning and cross-condition context.The appendix also reports that DeepSeek-R1 takes roughly four times longer than the other open-weight models because it emits visible chain-of-thought.
  • A.7 1:N Pilot: Specification, Prompt, and Diagnostics: The 1:N pilot varies authority metadata across otherwise identical cards, then fits standardized logistic regression and dominance analysis to derive signal weights.The pilot reports McFadden R2 = 0.033, described as a small but real metadata effect.
  • A.8 Weight-Sensitivity of the Direction Result: Weight-sensitivity analyses leave the directional conclusion intact: boosted flips favor higher authority under uniform and single-dimension alternatives, while flipped results remain non-directional.Under derived weights, boosted flips favor higher authority 68.2%; under uniform weights, the corresponding share is 67.1%.

E Tier Ablation: gpt-4o-mini

The gpt-4o-mini ablation tests whether resistance is specific to frontier-tier models rather than closed-weight models generally. Its results place the smaller model closer to the open-weight cohort and preserve the aggregate headline result.

  • Tier ablation: The gpt-4o-mini ablation repeats the headline protocol with identical prompts, parser, conditions, 2,250 cells, and a 100% parse rate.Its average runtime is approximately 2.2 seconds per cell.
  • Tier ablation: The ablation supports frontier-tier post-training, rather than closed-weightness alone, as the interpretation of reduced authority bias.The mini model reduces flips under anti-authority instructions, unlike the backfire observed across all three frontier closed-weight models; the nine-model aggregate flip rate is 39.0% versus 39.2% for eight models.

G Additional Per-Model Detail

Additional analyses show that susceptibility differs across models and dimensions, that frontier models agree unusually often, and that behavioral bias cannot be reliably inferred from explicit justification language alone. Model-level explanations remain partly speculative because deployed training details are undocumented.

  • Per-model directional pull: Gemma 2 rarely flips under boosted metadata, but 74.3% of its boosted flips favor higher composite authority.This separates the frequency of flipping from the direction of the resulting choice.
  • Per-model boost-pick lifts: Boost-pick lifts range from 36.5% for Qwen 2.5 to 23.6% for gpt-5.4 and claude-sonnet-4-6, with the four least-susceptible models also showing the lowest lifts.Across eight models, the extremes align, but the sample is too small to establish a strict population-level relationship.
  • Pairwise selection agreement: The three frontier models are the only off-diagonal pairs exceeding 60% selection agreement, reaching 63.0% for gpt-5.4 and claude-sonnet-4-6.These pairs span OpenAI, Google DeepMind, and Anthropic.
  • Frontier backfire: A mild anti-authority prompt backfires across frontier models, while strong content-first instructions reduce bias across all eight models.Replication across three vendors makes a single-vendor artifact less likely, and the contrast suggests wording-specific failure rather than complete unsteerability.
  • Hypothesized drivers: The paper cannot directly attribute closed-versus-open differences to specific architecture or training choices because deployed post-training specifications are undocumented.Pretraining composition and alignment regime are offered as plausible drivers rather than established explanations.
  • Justification language patterns: Explicit authority mentions are an unreliable proxy for behavioral bias: models can reduce authority language without proportionally reducing authority-driven selections.The appendix attributes this say-do gap to surface-level effects of instruction tuning and RLHF relative to upstream selection behavior.
Loading 2609.00248v1…