Where Does Retrieval-Based Open-Ended Evaluation Fail? Automatic Taxonomy Induction from Long-Form Medical Answer Factuality Verification

AI summary

Overview

• Research area: Natural language processing, specifically retrieval-augmented (RAG) factuality evaluation for long-form medical text, error taxonomy construction, and LLM-as-judge diagnostics.
• Technical level: Advanced. The paper assumes familiarity with retrieve-then-verify pipelines, atomic claim decomposition, retrieval metrics, and LLM verifier prompting.
• Scope: A model-agnostic error taxonomy plus a large-scale empirical audit of why retrieval-based open-ended medical factuality verification fails, tested across 4 retrievers, 3 knowledge corpora, and 6 verifier LLMs on the MedExpert dataset and 4 adapted closed-ended datasets.

What This Paper Is About

The dominant way to check whether an LLM's medical answer contains hallucinations is to decompose the answer into atomic claims, retrieve evidence from an authoritative corpus, and ask an LLM verifier to judge each claim. This paper shows that pipeline collapses in truly open-ended medical settings and then builds a fine-grained error taxonomy to explain where and why, since aggregate metrics like F1 cannot localize the failure. The authors induce that taxonomy automatically with LLM-as-judge so the diagnosis scales beyond manual inspection, and they stress-test whether better models, retrievers, or corpora fix the problem.

Key Contributions

1. A multi-dimensional error taxonomy for retrieval-based open-ended factuality evaluation, covering retrieval-stage errors across five quality dimensions (entity match, clinical scope, trustworthiness, evidence polarity, and chunking issues) and verifier-reasoning errors across six consecutive steps (selection, interpretation, inference, calibration, consistency, and grounding). The taxonomy is automatically inducible and requires no gold answers or annotated gold evidence.
2. A large-scale empirical analysis across 6 verifier LLMs, 4 retrievers, and 3 knowledge corpora, demonstrating that model scaling, added reasoning effort, medical fine-tuning, and corpus expansion all fail to resolve the dominant failure patterns.
3. Actionable recommendations: task-oriented multi-dimensional assessment of retrieved evidence paired with intermediate-step assessment of verifier behavior reflects system capability more faithfully than aggregated end-to-end metrics.
4. A human-in-the-loop LLM-as-judge pipeline for labeling evidence quality and classifying verifier reasoning errors at scale, with the final codebooks saturated after 6 iterations and validated against human annotation with inter-annotator agreement of 0.86 on claim-evidence pairs and 0.8 on claim-reasoning traces.

Main Findings

• Open-ended verification collapses relative to closed-ended benchmarks. In Table 2, using the Qwen3 retriever over the MEDIC corpus with the GPT-5.4 verifier, MedExpert reaches precision 5.6, recall 17.0, and F1 8.4 under the Refuted-only mapping, versus F1 of 67.2 on SciFact, 72.5 on SciFact-Open, 49.7 on CovidFact, and 38.9 on HealthVer. The paper's text describes the open-ended F1 as 0.08 (and elsewhere 0.06) versus 0.7 to 0.8 on CovidFact and SciFact-style benchmarks.
• Better retrievers help but do not solve the problem. Qwen3 is the strongest retriever (recall 25.6, F1 35.2 in Table 4) and produces the fewest No Evidence labels (8.0%) while converting most into Supported (78.3%) in Table 5, yet MedCPT and the hybrid RRF retrievers still leave recall at 19.8 (MedCPT), 24.4 (RRF-2), and 18.6 (RRF-4).
• Expanding to live web search does not help. Replacing MEDIC with Google General Search, Google Scholar Search, or the merged Google Search corpus drops recall and F1 in every case (Table 6: MEDIC 25.6/35.2; Google General 17.4/26.1; Google Scholar 15.1/23.9; merged Google Search 15.1/23.4). The authors note authoritative sources such as uclahealth.org can themselves contain unfactual content.
• Verifier capability does not follow benchmark ranking. Mistral Small 3 (24B) achieves the highest recall at 34.9 (F1 42.6) among the six verifiers, while GPT-5.4 reaches 25.6/35.2 and Gemma3-27B only 8.1/14.1 (Table 3). Higher healthcare leaderboard rank did not predict better open-ended verification performance.
• More test-time reasoning wastes tokens without improving recall. Increasing GPT-5.4's reasoning effort from Low to High raised reasoning tokens from 57,398 to 466,052 while recall stayed at 25.6 and F1 moved only from 35.2 to 37.3 (Table 7). The authors attribute this to hedging rather than better evidence use.
• Medical fine-tuning yields limited gains. MedGemma-27B roughly doubles Gemma3-27B's recall (16.3 versus 8.1) but remains far behind the strongest verifier, suggesting fine-tuning cannot exceed the base architecture's and pretraining's capabilities.
• Intermediate reasoning quality diverges from end-to-end rank. Counting a trace correct only when all 6 steps are correct, GPT-5.4 is highest at 90% and Mistral Small 3 lowest at 66%, implying Mistral Small 3's highest recall is largely a lucky guess based on faulty reasoning.
• Retrieved evidence is mostly low quality. Qwen3's high-quality-passage rate across all five retrieval dimensions is 14.5%, seven times MedCPT's 2%. Qwen3's remaining weaknesses are finding the correct clinical scope (27% exact scope match) and locating direct signals of affirmation or contradiction.
• Three structural properties drive the gap. The authors identify comprehensiveness of the evidence base, identifiability of gold evidence, and alignment between retrieved evidence and the claim. Roughly 90% to 81% of SciFact and SciFact-Open gold passages are contained in MEDIC, whereas COVID-related gold passages drawn from news, opinion, and lay sources fall outside it.
• End-to-end metrics alone can mislead model selection. The paper argues precision, recall, and F1 without multi-dimensional intermediate evaluation can point researchers to the wrong system.

Methodology in Plain English

The authors take an existing decompose-then-verify pipeline (MedScore, built on the FActScore idea) and instrument it. Each long-form medical answer is split into atomic, condition-aware claims; each claim is used as a retrieval query against the top 10 passages; and an LLM verifier assigns one of four labels—Supported, Refuted, No Evidence, or Ambiguous—instead of the usual binary verdict. That richer label set is then mapped back to binary True/False in three different ways to compute precision, recall, and F1 against clinician annotations.

To find out why the pipeline disagrees with clinicians, the team first wrote a seed error codebook from their interactions with clinicians, then ran an LLM-as-judge (Claude Opus 4.7) over the disagreement cases to surface additional error patterns. They iterated the codebook 6 times until it saturated—new data stopped producing new patterns. The final judge labeled sub-codes over 800 (claim, evidence) pairs for retrieval and 576 (claim, reasoning) traces for verification, all drawn from the MedExpert-gold subset. To check the judge, humans manually annotated 50 claim-evidence pairs and 60 claim-reasoning traces, yielding inter-annotator agreement of 0.86 and 0.8 respectively.

They then ablated the pipeline along three axes to test whether the errors were artifacts of old models: swapping among 4 retrievers (MedCPT, RRF-2, RRF-4, and Qwen3-Embedding-8B), swapping among 3 corpora (the 30.3M-passage MEDIC corpus, Google General Search, and Google Scholar Search), and swapping among 6 verifier LLMs. They also varied GPT-5.4's reasoning effort and compared medically fine-tuned versus general-domain open-source models of the same architecture.

The MedExpert-gold subset was constructed by having five clinicians adjudicate all 225 sentences originally annotated as factually incorrect; 86 were retained as containing factual errors and 139 were reclassified as correct. The authors note that recall remains generalizable to the full MedExpert set because all clinician-flagged "Has Error" sentences are included, while precision and F1 are not.

Why This Matters

Impact on research. The paper argues that aggregate F1 hides where retrieve-then-verify pipelines break, and that failure modes in open-ended medical verification are structural rather than a symptom of outdated models. It provides a reusable, annotation-free taxonomy and an automatic labeling pipeline for diagnosing them, and it challenges the common assumption that stronger retrievers, larger verifiers, or more test-time reasoning will fix medical factuality evaluation.

Real-world applications:

• Patient-facing medical question answering, where hallucinated claims about dosage, contraindications, or subpopulations must be caught before reaching patients.
• Clinical report generation and summarization, where fabricated findings can propagate into downstream care decisions.
• Automated fact-checking of health content at scale, including claims circulating in search results and lay media, which the paper shows often lack supporting evidence in scientific corpora.
• Model and vendor selection for healthcare deployments, where the paper shows leaderboard rank does not predict open-ended verification reliability.

Industry relevance. Any organization shipping RAG-based medical QA, clinical documentation, or health content moderation needs a way to audit its own pipeline rather than trusting a single F1 number. The taxonomy gives teams a checklist of evidence-quality and verifier-reasoning failure modes, and the finding that authoritative web sources can themselves contain falsehoods argues against substituting corpus expansion for verification quality.

Future Directions

• Build task-oriented retrieval evaluation and re-ranking. The authors recommend customized multi-dimensional assessment of retrieved passages instead of formula-based methods such as lexical or cosine similarity and Reciprocal Rank Fusion.
• Shift verifier evaluation to intermediate behavior. Instead of competing on precision/recall/F1, the paper calls for checking each of the six reasoning steps to confirm verifiers are reasoning correctly rather than guessing correctly.
• Add temporal and metadata dimensions. Because MedRAG provides no passage metadata, the authors could not evaluate publication date and omitted a Temporal dimension; they note it can be added if richer resources become available.
• Close the gold-annotation gap. The paper could not compute precision@k or recall@k for retrieval, or generator metrics as in RAGChecker, because open-ended tasks lack gold passages and gold answers. The authors note that if such resources are released, researchers can check whether these formula-based metrics align with the qualitative findings.
• Scale the automatic tagging beyond MedExpert-gold. Percentage estimates are restricted to the MedExpert-gold subset due to API budget limits and would be more comprehensive with more tagged data.

Target Audience

Researchers and practitioners working on retrieval-augmented generation, hallucination detection, and clinical NLP; engineers building or auditing medical QA and clinical documentation systems; and evaluation-focused scientists interested in LLM-as-judge pipelines, error taxonomy induction, and the limits of aggregate factuality metrics. Readers should be comfortable with retrieval pipelines (BM25, dense retrieval, rank fusion), atomic claim decomposition, and LLM verifier prompting.

Authors’ abstract

Retrieval-based factuality evaluation, where LLM-generated claims are verified against evidence from authoritative medical corpora, has become the dominant paradigm for scalable hallucination detection in high-stakes clinical settings. Despite the urgency of reliable and transparent medical fact verification, most systems measure performance with aggregate metrics like F1, which obscure where and why failures occur. Existing RAG diagnostics require gold answers or annotated gold evidence, neither of which exists in this regime. We introduce two comprehensive taxonomies, grounded in a case study on the open-ended MedExpert dataset and 3 closed-ended datasets, decomposing failures into retrieval-stage errors along five quality dimensions, and verifier-reasoning errors into six consecutive steps. We adapt an automatic pattern induction pipeline using LLM-as-Judge to label evidence quality and classify verifier reasoning errors at scale, and then stress-test our findings across 4 retrieval methods and 6 frontier verifier models. Our analysis reveals that scaling model size, adding reasoning effort, expanding to authoritative web sources, and applying medical fine-tuning do not resolve these failure modes, demonstrating that they represent fundamental limitations of the retrieve-then-verify paradigm in open-ended medical settings rather than artifacts of outdated systems. We release our code and data at https://anonymous.4open.science/r/Medical_RAG_eval-4AB5 for the full reproducibility of our results.