PaperlayerView exact revision on arXiv

Source-linked AI summary

Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence

Owen Walters, Lihu Chen, Francesca Toni

arXiv:2610.21731v1cs.CLcs.AIcs.IR

TL;DR

Search-augmented LLMs can answer from outdated knowledge even when relevant updated evidence is available, leaving unclear why they fail to use it. The paper evaluates deployed and controlled search with post-cutoff facts, finding snippet reliance associated with stale answers and unfaithful citations. Its layer-wise analysis finds that retrieved evidence can influence internal predictions without winning in the final answer.

  • Problem

    It is unclear why search-augmented LLMs sometimes fail to use relevant updated evidence, despite its availability.

    Source: §1 · p. 1

  • Method

    The authors evaluate a proprietary model under deployed search and it plus six open-weight configurations under controlled retrieval, using post-cutoff facts and layer-wise analysis.

    Source: §1 · p. 2, §5 · p. 4

  • Results

    Most stale answers and unfaithful citations occur in snippet-only episodes, and layer-wise readouts show retrieved answers may gain an early advantage yet lose to outdated answers in final layers.

    Source: §1 · p. 2, §4.1 · p. 3, §5 · p. 4

  • Takeaways & Limitations

    Providing source pages directly reduces some failures but does not eliminate them; the findings may help guide development of search-augmented LLMs that use evidence better and respond more reliably.

    Source: §6 · pp. 4–5

    GLM-5.3-Flash discloses no training cutoff, so the authors check for contamination through its behavior instead.

    Source: Limitations · p. 5

Abstract

from arXiv · show

Web search allows LLMs to access up-to-date information, but it remains unclear why they sometimes produce outdated answers even when relevant evidence is available. To study this problem, we construct temporal facts whose values changed after the models' training cutoffs. We evaluate one proprietary model under deployed web search and six open-weight model configurations under controlled retrieval. We find that models often answer from search snippets without opening the source pages, which is associated with higher rates of outdated answers and unfaithful citations. Our controlled experiments and layer-wise analysis reveal a gap between the availability of evidence and its effective use. Even when retrieved evidence influences internal predictions, models may still favour outdated information in their final answers. We hope these findings will help advance research on the reliable use of retrieved evidence in search-augmented LLMs.

1 Introduction

The paper asks why search-augmented LLMs fail to use relevant evidence and investigates this with post-cutoff facts, deployed search, controlled retrieval, and mechanistic analysis. Models often rely on snippets without opening pages, and retrieved evidence may influence internal predictions without displacing stale answers.

1 Introduction

  • The paper investigates why search-augmented LLMs can give outdated answers or cite contradictory sources despite relevant updated evidence being available.The authors note that citations lend answers provenance.Source: §1 · p. 1
  • The authors construct 225 post-cutoff temporal facts across six domains, each with an outdated and a current value, to assess evidence use and citation support.Facts are generated from primary sources with a deterministic, regenerable pipeline.Source: §1 · p. 1
  • The study first evaluates a proprietary model with native web search, then tests recurrence across that model and six open-weight configurations given identical frozen search results.The analysis also uses controlled experiments and layer-wise readouts to examine remaining failures.Source: §1 · pp. 1–2
  • Models often answer from snippets without opening source pages; most stale answers and unfaithful citations occur in these episodes, while retrieved evidence can still lose to stale information in final answers.Providing source pages directly reduces some errors but does not eliminate them.Source: §1 · p. 2

2 Related Work

Prior work has measured temporal answer accuracy, search use, citation faithfulness, and conflicts between model knowledge and external evidence. This paper focuses on whether web-search evidence is actually used, across live and controlled search settings.

Temporal knowledge and web search

  • Earlier temporal benchmarks assess correctness, while Kale (2026) reports frequent native-search use without checking whether cited pages support answers; this paper examines snippet reliance across proprietary and open-weight models.Other search-agent evaluations also report premature stopping and top-result reliance (Gupta et al., 2026; Zhang et al., 2025).Source: §2 · p. 2

Citation faithfulness

  • Prior citation work distinguishes source support from whether a source produced the answer; this study uses post-cutoff facts to examine citation faithfulness under live and controlled web search.Related studies examine conflicts between parametric knowledge and external evidence through external behaviour and internal representations.Source: §2 · p. 2

3 Setup

The study builds a controlled evaluation around post-cutoff facts, multiple search settings, and varied prompts. It tracks how deeply models read, answer accuracy, and whether citations contradict asserted values.

3 Setup

  • The dataset includes 225 treatment facts across six domains, excludes changes effective by 16 February 2026 or announced before 1 February, and uses deterministic extraction.FDA claims answerable from standard medical knowledge are excluded; the model-assisted FDA clause rewriting is manually reviewed.Source: §3 · p. 2
  • The study compares Luna’s native web search with controlled retrieval for Luna and six open-weight configurations, using the same ten-result menu for each of 223 facts.The ground-truth page is reachable in the menu, and search, open-page, and find tools are shared.Source: §3 · p. 2
  • Prompts include bare and source-requested open questions, claim checks with current or outdated values, and four further types; every prompt states the date.The additional types are cued open questions, quote requests, citation requests, and fact-sheet checks.Source: §3 · p. 2
  • Reading depth distinguishes snippet-only episodes, opened-and-cited pages, and opened-but-uncited pages; accuracy excludes abstentions, and citations are unfaithful when pages contradict asserted values.The caption defines Table 1’s accuracy, snippet, unfaithful-citation, and control-fact-floor columns as percentages.Source: §3 · pp. 2–3, Table 1 · p. 3

4 Results

Search-augmented models often answer from snippets without opening pages, and those episodes are associated with more stale answers and unfaithful citations. Direct page access and prompt restatement reduce some failures, while similar failure patterns appear across models.

4.1 Search-Augmented LLMs Often Rely on Snippets

  • Search raises claim-check accuracy by 27–42 points, while 2.6–76.0% of episodes involve answering without opening any page.Under native search, Luna answers 53.6% of episodes from snippets alone; open-question accuracy reaches 85.8–95.4%.Source: §4.1 · p. 3
  • Across seven configurations, snippet-only episodes have higher stale-answer and unfaithful-citation rates than episodes where an opened page is cited.The respective rates are 6.4% versus 1.3% for stale answers and 6.4% versus 2.3% for unfaithful citations.Source: Table 2 · p. 3, §4.1 · p. 3

4.2 Failures Are Shared across Models

  • Seven of nine failure types under native search are positively associated with failures under frozen retrieval (odds ratios 1.84–6.46, p < 0.05).A negative control shows no association (odds ratio 1.05).Source: §4.2 · p. 3

4.3 Direct Access to Evidence Reduces Failures

  • Providing the same Wikipedia page in the prompt reduces stale-claim endorsements for gemma-4-31b and gpt-oss-20b, while prompt restatement recovers 41 of 157 failed fact-sheet checks.qwen3-8b benefits less from direct page access, improving from 16/49 to 8/49 stale endorsements.Source: §4.3 · p. 3

5 Mechanistic Analysis of Evidence Use

Layer-wise analysis finds that retrieved evidence can influence predictions even when a model ultimately favors an outdated answer. The analysis compares answer scores across layers with and without the source page.

5 Mechanistic Analysis of Evidence Use

  • Direct evidence reduces failures but does not eliminate outdated answers, motivating analysis of how models process the available evidence.Source: §5 · p. 4
  • The analysis compares prompts asserting an outdated club with and without the player’s current Wikipedia page, tracking candidate-club scores across layers.The logit lens and Jacobian lens project intermediate hidden states at the answer position onto token-level logits.Source: §5 · p. 4
  • In failures, the retrieved value gains support in upper layers but the outdated value can overtake it in final layers; gemma-4-31b does not suppress the outdated value.Across models, the page promotes the retrieved value in failures, though less than in updates (∆new > 0, p ≤ 0.04).Source: §5 · p. 4

6 Conclusion

The conclusion emphasizes that retrieved evidence can be available and influence internal predictions without determining the final answer. The findings may help guide development of more reliable search-augmented responses.

6 Conclusion

  • Models often answer from snippets without opening source pages, and those answers are more likely to be stale and carry unfaithful citations.Source: §6 · p. 4
  • Retrieved answers can gain an early advantage in upper layers but lose to outdated answers in the final layers.Source: §6 · p. 4
  • The findings may help guide search-augmented LLMs toward better use of retrieved evidence and more reliable responses.Source: §6 · p. 5

Limitations

The study’s evidence is bounded by a single proprietary model and domain constraints; fixed answer formatting and rare labels also qualify its measured failure rates.

Limitations

  • Only one proprietary model—the cost-selected lightest GPT-5.6 tier—is studied, and additional frontier models may behave differently.Source: Limitations · p. 5
  • The fixed answer-line contract may change behavior versus free prose; it lowered Luna’s false endorsement by 17.5 points, making deployed failures likely lower bounds.Source: Limitations · p. 5
  • Unfaithful citations are rare for Luna (1.0%) and GLM-5.3-Flash (1.8%), making rates sensitive to grading errors; humans assign or confirm every adverse label requiring judgment.Source: Limitations · p. 5
  • Failure rates vary by domain: sports drives severe rates, while SEC (13 facts) and chemical (8) are underpowered because few usable post-cutoff facts exist.Source: Limitations · p. 5
  • Layer-wise analysis uses sports transfers because club names are short, characterizing the failure without causally testing a mechanism.Source: Limitations · p. 5

Ethical Considerations

The appendix documents the study’s ethical safeguards and the construction, evaluation, and analysis procedures behind its results. The experiments use public temporal facts, controlled retrieval, and multi-stage grading, with explicit cautions about medical use and misuse of snippets.

Ethical Considerations

  • The study reports only aggregate failure rates, releases no snippet-manipulation tools, uses public facts, and warns that its drug-label facts must not be used medically.The people facts concern named individuals and cannot be anonymised; no human participants were involved beyond the authors’ grading.Source: Ethical Considerations · p. 5

A Models

  • The evaluation studies Luna under native web search and Luna plus six open-weight configurations under frozen retrieval; three open models also support layer-wise readouts.Frozen retrieval presents a shared menu of ten results, with page-opening and turn limits.Source: A Models · p. 7

B Fact Sources

  • The fact set contains 225 post-cutoff changes across six domains, derived from primary sources and filtered for reliable ground truth and model cutoffs.FDA changes are identified by comparing current label clauses with archived pre-cutoff versions; controls use unchanged facts.Source: B Fact Sources · p. 7

C Prompt Types

  • The study pairs open questions and claim checks with additional cued questions, quote requests, citation requests, and fact-sheet checks for pooled analyses.Some frozen-retrieval prompts reuse evidence from a wording or claim twin; adding a required answer line reduced Luna’s closed-book false endorsements by 17.5 points.Source: C Prompt Types · pp. 7, 9, Table 5 · p. 8

D Grading

  • Grading uses deterministic checks first, then a calibrated LLM marker for passing outcomes, with humans deciding adverse or unresolved cases and auditing passes.The authors hand-graded 4,108 answers on 3,856 cards and report blind consistency checks.Source: D Grading · p. 9

E Accuracy with and without Search

  • Search raises claim-check accuracy by 27.0 points for qwen3-8b to 42.1 for gpt-oss-120b, while closed-book accuracy lies in a narrow band.Accuracy excludes abstentions and counts correct answers among correct, stale, or other answers.Source: E Accuracy with and without Search · p. 9, Table 7 · p. 10

F Snippet Content by Domain

  • Among open models, sports answers from snippets alone are stale in 20.2% of episodes (53/262), versus 2.1% (1/48) when an opened page is cited.Snippet-value coverage is assessed against the study’s grading criteria for each statement.Source: F Snippet Content by Domain · p. 9, Table 8 · p. 10

G Fact-Sheet Outcomes by Model; H Handed-Over versus Searched-For Pages

I Shared Failures across Models

The section compares Luna’s native-search outcomes with outcomes under frozen retrieval for the same prompts. It defines carry-over and new failures, and reports positive associations for several failure types while a negative control shows no association.

I Shared Failures across Models

  • The comparison joins outcomes for each prompt across settings; “carried” is the share of Luna-failed prompts that fail again, while “new” is the failure share among Luna-passed prompts.Table 11 reports odds ratios with 95% Woolf confidence intervals.Source: I Shared Failures across Models · pp. 10–11, Table 11 · p. 11
  • For stale answers joined on the route matching the answer value, pooling over all three routes yielded an odds ratio of 2.66, with an interval of 1.42–4.96.The negative control occurred at the same rate whether or not Luna failed the prompt, so the join did not produce spurious carry-over.Source: I Shared Failures across Models · p. 11
  • For stale answers joined on the answer-value-matching route covering sports, finance, SEC, and chemical facts, the pooled odds ratio was 2.66 [1.42, 4.96]; people and FDA routes had no carry-overs (0/18; 0/12).For citing a URL absent from the model’s own search session, the interval spans 1 (1.93 [0.97, 3.82]).Source: I Shared Failures across Models · p. 11

J Layer-Wise Readout Details

The layer-wise analysis compares how retrieved and asserted club representations shift when the page is added. Retrieved clubs rise in failures too, but suppression of the asserted club is inconsistent across models and readouts.

J Layer-Wise Readout Details

  • The two metrics compare each club’s answer-slot score with and without the page, averaged over the upper third of layers; the prompts differ only in page content.The paper defines the readout as a club’s score under a lens at each layer.Source: Table 12 · p. 11, J Layer-Wise Readout Details · p. 11
  • Club scores use candidate first tokens from club names, taking the maximum score and minimum rank; token-set misses occur in roughly a third to a half of replies.At the penultimate layer, readout answers agree with the model’s answer on 85.1% of episodes under the logit lens and 85.6% under the J-lens.Source: J Layer-Wise Readout Details · p. 11
  • The layer averages use the upper third of each model’s stack and exclude the first and final layers; the J-lens is severely ill-conditioned.The authors compare magnitudes only within a model and use Mann–Whitney U tests between outcome groups.Source: J Layer-Wise Readout Details · pp. 11–12
  • In failures, the retrieved club’s score rises in every row (p ≤0.04), but asserted-club suppression is not significant for gemma or qwen under the J-lens.Suppression is significant for gpt-oss under both readouts and qwen under the logit lens; magnitudes should be compared only within a model.Source: Table 13 · p. 12
  • The penultimate-layer rank analysis reports median ranks for retrieved and stale clubs under the logit lens and the share of facts with the retrieved club in the top 15.The layer plot tracks both clubs’ median ranks through the stack.Source: Table 14 · p. 12, J Layer-Wise Readout Details · p. 12
  • For qwen3-8b on Anthony Gordon, adding the page raises Barcelona’s logit-lens score by 15.59 and lowers Newcastle United’s by 5.96 at layer 35.Barcelona gains 21.55 on Newcastle; with the page the model answers Barcelona, while without it the model answers Newcastle United. The same prompt leaves gpt-oss-20b answering Newcastle.Source: J Layer-Wise Readout Details · p. 12, Table 15 · p. 12

K Re-Asking after the Evidence

This experiment tests whether repeating the full prompt after each opened page helps models revisit failed fact-sheet checks. It holds the retrieval materials and sampling settings constant, and reports recovery counts before and after exclusions.

K Re-Asking after the Evidence

  • The rerun selects fact-sheet checks whose original reply did not name exactly the stale line, with or without a source request; unparseable verdicts count as failures.Each opened-page result is followed by the full prompt, making the question the most recent text read by the model.Source: K Re-Asking after the Evidence · p. 12
  • Restating the question after each opened page recovered 55 of 222 failed fact-sheet checks as run (24.8%), before the paper’s exclusions.The rerun retains 157 episodes after exclusions and keeps the menu, page bytes, sampler, models, and episode identifiers unchanged.Source: K Re-Asking after the Evidence · p. 12, Table 16 · p. 14
Loading 2610.21731v1…