cs.CLOct 16, 2026

Recognizing Study Lineage Is Not Enough: Over-Crediting Same-Trial Evidence in LLM-Based Medical Verification

Authors: Mukand Krishna, Anoushka Harit, Chenghao Xiao, Zhongtian Sun

Organizations: School of Computing, University of Kent, UK · Cancer Research UK Cambridge Institute, University of Cambridge, UK · Shanghai University of Finance and Economics, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Publication count is not study count. A model can identify shared trial lineage correctly in a direct task yet over-credit a second publication from the same trial when aggregating evidence. We evaluate this behavior in medical LLM verification using 14 source-audited clinical families and five model conditions. The families span provenance, clinical applicability, result direction, and composition; the provenance families provide the direct lineage tests. Recognition and aggregation are separate stateless calls, and repeated calls assess within-input stability. Across all three original provenance families, GPT-5.5 identifies the independent update in 10/10 recognition calls per family but over-credits the same-trial report in 5/5 separate aggregation calls per family. Selected Qwen3-4B and Ministral-3-8B conditions show the same cross-task contrast. A study-grouping instruction reduces GPT-5.5 over-credit from 4/5 to 0/5 in each later pure-provenance family. Other model conditions show mixed behavior, and rejected outputs limit several comparisons. The results identify a behavioral gap between elicited lineage recognition and evidence aggregation. The code and data artifacts are available at https://anonymous.4open.science/r/study-lineage-review-artifact.

Figures & tables

Figure 1: A source-audited example of the recognition–aggregation gap. The STRIVE primary report and later analysis derive from trial NCT02456740, whereas the ARISE primary report comes from an independent trial (NCT02483585) ( Goadsby et al., 2017 ; Diener et al., 2021 ; Dodick et al., 2018 ) . In separate stateless GPT-5.5 calls, ARISE is identified as the independent update in 10/10 recognition calls (five per candidate order), but the later STRIVE analysis is marked as independent corroboration in 5/5 aggregation calls. Repeated calls assess within-input stability; the clinical family is the experimental unit.
FamilyClinical targetTested contrastInputs
CF-P1Erenumab; episodic migraineSTRIVE post-hoc analysis vs ARISE6
CF-P2Dupilumab; uncontrolled asthmaQUEST analysis vs distinct trial6
CF-P3SGLT2 inhibition; kidney diseaseEMPA-KIDNEY analysis vs distinct trial6
CF-A1Adjuvant pembrolizumab; melanomaapplicable vs out-of-scope evidence6
CF-A2Factor Xa inhibition; atrial fibrillationapplicable vs out-of-scope evidence6
CF-A3Olaparib; ovarian-cancer maintenancefirst-line vs relapsed disease6
Table 1: Clinical targets and analytical roles of the selected 14 families. Inputs are distinct task inputs per model, each invoked five times. The first ten families belong to the original suite (66 inputs); the final four were frozen later (44 inputs). Source identifiers and publication anchors appear in Appendix Tables 4 and 5 .
ModelRecognitionAggregationRejected R/A
GPT-5.53/30/30/0
Qwen3-4B3/30/20/5
Gemma3-4B0/3N/E0/10
Llama-3.1-8B0/3N/E0/10
Ministral-3-8B1/30/110/5
Table 2: Family-level recognition–aggregation contrast in CF-P1–P3. Full recognition requires 10/10 correct calls across both counterbalanced orders. The aggregation column counts lineage-respecting families among fully recognized families with valid same-trial aggregation outputs. N/E means not evaluable because no family met this prerequisite. A family is lineage-respecting when the added same-trial report is not labeled as independent corroboration. R/A denotes rejected recognition/aggregation calls.
Figure 2: Model-by-family results for lineage recognition and same-trial evidence aggregation. Panel A reports the original provenance families; Panel B reports the later pure-provenance families used to test study grouping. R denotes full recognition across both candidate orders (10/10 correct calls); D and G indicate whether document-level and study-grouped aggregation, respectively, avoided same-trial over-credit in every valid call. A blue check marks that the criterion was met; an orange cross indicates over-credit in at least one valid call. A triangle marks order-sensitive recognition, a question mark marks recognition limited by rejected outputs, and a dash indicates that aggregation is not estimable. Preservation of genuinely independent evidence is evaluated separately in Appendix Table 8 .
Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Design componentFamilies/conditionsAnalysis statusRole in the paper
Original suiteCF-P1–P3, CF-A1–A3, CF-D1–D3, H3prospectively frozen primary evaluationrecognition and aggregation across provenance, applicability, direction, and composition
Later packageP5R, P6R, H5, H6separately frozen follow-upstudy grouping and preservation of genuinely independent evidence
Prompt-order auditP5R, P6Rsecondary robustnesstests one semantically equivalent verifier-prompt reordering
Oracles and cue ablationCF-P3 oracles, H3-M1post-hoc diagnostictests explicit relation/policy information and removal of the study-group cue
Lineage ×\times overlapProMISe/ARISE controlexploratory, policy-instructedseparates application of a stated lineage rule from ordinary passage overlap
Appendix
Table 3: Experimental chronology and analysis status. Later diagnostics were not part of the original primary denominator and are reported separately.
FamilyInputsTested claim/contrastStudy identifiers recorded in selected verifier inputs
CF-P16Erenumab and episodic-migraine frequencySTRIVE: NCT02456740 ; ARISE: NCT02483585
CF-P26Add-on dupilumab in uncontrolled asthmaNCT02414854 ; NCT01854047
CF-P36SGLT2 inhibition in chronic kidney diseaseNCT03594110 ; NCT03036150
CF-A16Adjuvant pembrolizumab and melanoma recurrenceNCT02362594 ; S1404; NCT01866319
CF-A26Factor Xa inhibition for atrial-fibrillation stroke preventionNCT00412984 ; NCT00403767 ; NCT00643201
CF-A36First-line olaparib-containing ovarian-cancer maintenanceNCT01844986 ; NCT02477644 ; NCT00753545
Appendix
Table 4: Selected family inventory. P, A, and D isolate provenance, applicability, and direction; H families combine relations. “Inputs” counts distinct task inputs per model, each invoked five times. The first ten rows are the original suite (66 inputs; 330 calls per model); the final four are the later package (44 inputs; 220 calls per model). Listed identifiers cover the selected verifier inputs, not necessarily every publication or every comparison-task source.
FamilyOne initial-source label in the frozen inputPMID
CF-P1A Controlled Trial of Erenumab for Episodic Migraine29171821
CF-P2Dupilumab Efficacy and Safety in Moderate-to-Severe Uncontrolled Asthma29782217
CF-P3Empagliflozin in Patients with Chronic Kidney Disease36331190
CF-A1Adjuvant pembrolizumab versus placebo in resected stage III melanoma33857412
CF-A2Apixaban versus warfarin in patients with atrial fibrillation21870978
CF-A3SOLO1/GOG 3004 maintenance olaparib in newly diagnosed advanced ovarian cancer34715071
Appendix
Table 5: One initial-source anchor per family. These metadata identify the archived test inputs; they do not substitute for checking every added publication, and the tested claim is not asserted here as a current clinical recommendation.
BlockPlanned calls per tested modelTested model conditionsPlanned calls
Original ten-family suite33051,650
Later four-family package22051,100
Prompt-order robustness403120
CF-P3 oracle + H3-M1 diagnostics653195
Exploratory lineage-by-overlap control–model-dependent36 total
Appendix
Table 6: Call accounting by analysis block. Completion does not imply validity; rejected outputs are reported separately in Appendix H .
ModelFamilyRecognition correct/validSame-trial over-credit/valid
GPT-5.5CF-P110/105/5
 CF-P210/105/5
 CF-P310/105/5
Qwen3-4BCF-P110/105/5
 CF-P210/100/0 (5 rejected)
 CF-P310/105/5
Appendix
Table 7: Recognition and same-trial aggregation for every model in CF-P1–P3. Rejected calls are shown explicitly; 0/0 has no valid semantic estimate.
FamilyModelRecognition correct/valid (rej.)D over-credit/valid (rej.)G over-credit/valid (rej.)
P5RGPT-5.510/10 (0)4/5 (0)0/5 (0)
 Qwen3-4B10/10 (0)5/5 (0)5/5 (0)
 Gemma3-4B5/10 (0)5/5 (0)5/5 (0)
 Llama-3.1-8B5/10 (0)0/5 (0)0/5 (0)
 Ministral-3-8B10/10 (0)0/0 (5)0/0 (5)
P6RGPT-5.510/10 (0)4/5 (0)0/5 (0)
Appendix
Table 8: Pure-provenance P5R/P6R results. Recognition and aggregation are separate stateless calls. Panel A counts same-trial over-credit among valid outputs; panel B counts support for the independent-trial update (flags/valid calls). In either panel, 0/00/0 means five rejected outputs, not a measured zero or loss of support.
FamilyModelRecognition correct/valid (rej.)D global flag/valid (rej.)G global flag/valid (rej.)
H5GPT-5.520/20 (0)5/5 (0)4/5 (0)
 Qwen3-4B15/20 (0)5/5 (0)5/5 (0)
 Gemma3-4B10/20 (0)0/0 (5)0/0 (5)
 Llama-3.1-8B10/20 (0)5/5 (0)0/0 (5)
 Ministral-3-8B20/20 (0)0/0 (5)0/0 (5)
H6GPT-5.510/10 (0)5/5 (0)1/5 (0)
Appendix
Table 9: Compositional H5/H6 results. D and G count the global independent_corroboration flag on a same-trial update, not attributable over-credit: both baselines contain an independent study. H5 recognition includes provenance and applicability comparisons. A 0/00/0 cell has no valid semantic estimate.
ModelCF-A1CF-A2CF-A3CF-D1CF-D2CF-D3H3
GPT-5.510/1010/1010/1010/1010/1010/1030/30
Qwen3-4B10/1010/1010/100/5†0/5^{\dagger}5/1010/1030/30
Gemma3-4B5/105/1010/105/105/5†5/5^{\dagger}5/1015/30
Llama-3.1-8B5/100/5†0/5^{\dagger}5/105/105/105/1015/30
Ministral-3-8B10/1010/1010/1010/105/1010/1020/30
Appendix
Table 10: Comparative recognition in applicability (A), direction (D), and compositional (H3) families. Cells are correct/valid calls. Each primitive family has ten planned recognition calls; H3 has 30 across three relations. †\dagger marks five additional rejected recognition calls, not five incorrect answers. Repetitions are within-input stability checks.
ControlFamilies/conditionsExpected handlingGPT-5.5 result
Exact repeat / nullCF-P1–P3, CF-A1–A3, CF-D1–D3no new independent evidenceduplicate or unchanged in 45/45 calls
Same-trial updateCF-P1–P3no independent corroborationover-credit in 15/15 calls
Independent updateP5R, P6Rpreserve independent support5/5 →\to 5/5 in each family
Out-of-scope positiveCF-A1–A3ignore for the fixed claimout_of_scope and ignored, 15/15
Challenging updateCF-D1–D3retain support and register conflictconflict or mixed evidence, 15/15
Appendix
Table 11: Compact control-outcome summary for the approved confirmatory and later pure-provenance conditions. The 45 exact-repeat calls are nine primitive families with five repetitions each; the 15/15 same-trial, out-of-scope, and challenging-update entries are three families with five repetitions each. The independent-support row is the separate P5R/P6R safety check.
FamilyTarget updateGPT-5.5 integration outcome
CF-A1applicable vs positive out of scopeapplicable update used; out-of-scope update ignored, 5/5
CF-A2applicable vs positive out of scopeapplicable update used; out-of-scope update ignored, 5/5
CF-A3first-line vs relapsed maintenanceapplicable update used; out-of-scope update marked out_of_scope and ignored, 5/5
CF-D1agreeing vs mortality-challenging ICU evidencechallenging update treated as conflict/mixed, 5/5
CF-D2agreeing vs mortality-challenging diabetes evidencechallenging update treated as conflict/mixed, 5/5
CF-D3agreeing vs annual high-dose vitamin-D evidenceagreeing update supports 5/5; challenging update conflicts and yields mixed_evidence, 5/5
Appendix
Table 12: Family-level applicability and direction integration for GPT-5.5. These are separate neutral-verifier calls from the comparative-recognition calls; the table is a control-axis audit, not evidence that applicability or direction is universally solved.
ModelSame-trial over-creditIndependent audit passStrict validCanonical valid
GPT-5.516/2014/2040/4040/40
Qwen3-4B20/2015/2040/4040/40
Gemma3-4B20/2020/200/4040/40
Appendix
Table 13: Secondary prompt-order robustness audit on P5R/P6R. Each model has 40 calls across two prompt variants, two families, same-trial and independent updates, and five repetitions. Same-trial over-credit is a forbidden independent_corroboration flag; the independent audit requires both the flag and use of the added independent evidence. Gemma’s canonical validity reflects removal of one outer Markdown JSON fence; raw outputs remain unchanged.
CellSource passageTokensCharsJaccardROUGE-LCosine
AnchorProMISe primary numerical result68335–––
Same/highProMISe HTA numerical result572990.6346150.4000000.846106
Same/lowProMISe HTA abstract conclusion301820.1093750.1224490.656978
Independent/highARISE primary numerical result563030.4237290.3709680.922640
Independent/lowARISE primary conclusion231320.1186440.1098900.678653
Appendix
Table 14: Frozen overlap measurements for the one-family policy-instructed lineage control. Scores are computed against the initial ProMISe numerical-result anchor; “high” and “low” are within-lineage strata, not causal treatments.
CellExpected relationDuplicate evidenceIndependent corroboration
Same/highsame trial5/50/5
Same/lowsame trial5/50/5
Independent/highindependent trial0/55/5
Independent/lowindependent trial0/55/5
Appendix
Table 15: Policy-defined lineage-by-overlap results for GPT-5.5. All 20 calls labeled the claim supported. The expected pattern is duplicate evidence for same-trial passages and independent corroboration for independent-trial passages. The four cells each contain five fresh stateless calls. This one-family result reports policy application across overlap strata. Separate designs are needed to estimate overlap effects or unprompted lineage discovery.
 CF-P3 oraclesH3-M1 cue ablation
ModelRelation: samePolicy: sameIndependent R/PRecognitionSame-trial flag
GPT-5.50/50/55/5, 5/510/105/5
Qwen3-4B5/55/55/5, 5/55/105/5
Gemma3-4B0/0 (5 rej.)0/0 (5 rej.)5/5, 5/55/105/5
Appendix
Table 16: Post hoc diagnostic outcomes, not additional primary families. R/P denotes the independent-trial condition under the relation and policy oracles, respectively. The two CF-P3 “same” columns and the H3-M1 same-trial column count the independent-corroboration flag on a same-trial update. A 0/00/0 cell has five rejected outputs and no semantic estimate. Qwen has five further rejected H3-M1 exact-repeat outputs; Gemma has 25 rejected diagnostic outputs overall. GPT-5.5 has no archived error row in this block.
FamilyPlannedGPT-5.5Qwen3-4BGemma3-4BLlama-3.1-8BMinistral-3-8B
CF-P13030/030/015/1520/1030/0
CF-P23030/010/2010/2020/1010/20
CF-P33030/025/510/2025/525/5
CF-A13030/030/025/525/530/0
CF-A23030/030/020/1025/530/0
CF-A33030/020/1015/1510/2020/10
Appendix
Table 17: Canonical-valid/rejected calls in the selected 14-family comparison. Cells record response-contract validity rather than correctness; rejected calls receive no semantic score. Planned calls are per model ; each cell sums to its row’s planned count.
More in this topic

LLM Reliability

Large language model reliability studies whether outputs remain correct, calibrated, robust, and controllable across prompts, contexts, and deployment conditions.

See all 759 papers

Explore similar work

CardsList

  1. Med-V1: Small Language Models for Zero-shot and Scalable Biomedical Evidence Attribution

    Mar 5, 2026Qiao Jin, Yin Fang, Lauren He +12LLM Answer VerificationHealthcare

    Division of Intramural Research, National Library of Medicine, National Institutes of Health · Department of Computer Science, University of Virginia · Center for Cancer Research, National Cancer Institute, National Institutes of Health +3

  2. The Provenance Gap in Clinical AI: Evidence-Traceable Temporal Knowledge Graphs for Rare Disease Reasoning

    Apr 18, 2026Md Shamim Ahmed, Maja Dusanic, Moritz Nikolai Kirschner +4Data ProvenanceHealthcare

    Department of Mathematics and Computer Science, University of Southern Denmark, Campusvej 55, 5230 Odense, Denmark · Universitätsmedizin Göttingen, Robert-Koch-Straße 40, 37075 Göttingen, Germany

  3. ProvenanceGuard: Source-Aware Factuality Verification for MCP-Based LLM Agents

    Jun 16, 2026Ander Alvarez, Santhiya Rajan, Samuel Mugel +1LLM Answer VerificationData Provenance

    1Multiverse Computing, Parque Cientifico y Tecnol´ogico de Gipuzkoa, Paseo de Miram´on, 170, 20014 Donostia / San Sebasti´an, Spain · 2Multiverse Computing, Centre for Social Innovation, 192 Spadina Avenue Suite 509, Toronto, ON M5T 2C2, Canada · 3Donostia International Physics Center, Paseo Manuel de Lardizabal 4, E-20018 San Sebasti´an, Spain +1