MedRAGChecker: Biomedical Claim Verification Framework

• MedRAGChecker is an evidence verification framework that decomposes biomedical responses into discrete, atomic claims for precise fact-checking.
• It integrates natural language inference and knowledge graph consistency checks to assess claim accuracy and identify hallucinations.
• Empirical evaluations reveal reduced hallucination rates and enhanced diagnostic metrics, supporting its use in high-stakes medical QA tasks.

MedRAGChecker is a claim-level, evidence verification and diagnostic framework for biomedical retrieval-augmented generation (RAG) systems, designed to scrutinize long-form LLM responses for factual support, hallucination, contradiction, and provenance at the atomic claim level. It operates by decomposing LLM-generated biomedical answers into discrete claims, checking each claim's evidentiary grounding using both natural language inference (NLI) mechanisms and biomedical knowledge graph (KG) consistency signals, and aggregating these results into robust answer-level diagnostics applicable to high-stakes medical question answering tasks (Ji et al., 10 Jan 2026).

1. System Architecture and Workflow

The MedRAGChecker pipeline processes a tuple $(q, D, a)$, where $q$ is a biomedical question, $D = \{d_1, \ldots, d_k\}$ is a set of top-$k$ retrieved evidence passages, and $a$ is a RAG-generated long-form answer. The pipeline consists of four principal stages:

1. Claim Extraction: The answer $a$ is parsed into a sequence of atomic factual claims $C = \{c_1, \ldots, c_n\}$, typically in subject–relation–object (SRO) triple form.
2. Textual Verification (NLI): For each $c_i$, an NLI verifier evaluates support against the evidence $D$, yielding a discrete verdict $$\hat{y}_i \in \{\textrm{Entail}, \textrm{Neutral}, \textrm{Contradict}\}$$ and an entailment probability $q$0.
3. KG-based Consistency: Each $q$1 is mapped to candidate knowledge graph triples in DRKG, and a plausibility score is computed using embedding-based and string-alignment methods.
4. Signal Fusion and Aggregation: Textual and KG support scores are combined into a calibrated support score $q$2 for each claim, and summary diagnostics (faithfulness, hallucination rate, context precision, safety error rate) are computed for the overall answer.

This modular design enables fine-grained dissection of RAG failures, distinguishing between unsupported factual claims, contradictions, and retrieval omissions.

2. Claim Extraction and Representation

The extraction process decomposes long-form text $q$3 into a set of atomic claims, each representing a verifiable biomedical statement:

• Teacher Extraction: A teacher LLM (GPT-4.1) is prompted with $q$4 and produces a JSON list of SRO triples, adhering to constraints of atomicity (no conjunction), faithfulness (no hallucinated content), and explicit handling of negation and uncertainty.
• Model Distillation: A compact biomedical LLM is fine-tuned on the teacher’s outputs using sequence-to-sequence cross-entropy:

$q$5

Distilled claim extractors (e.g., Med42-Llama3-8B) achieve 75–85% claim-level span F1 against teacher outputs on biomedical QA texts.

• Inference Efficiency: Extraction models are LoRA-adapted for scalability, enabling per-example inference in seconds and removing reliance on external API calls at test time.

3. Hybrid Verification Modules

3.1 Evidence-Grounded NLI

Each claim $q$6 is evaluated against $q$7 using an NLI classifier:

• Annotation: The teacher model labels each $q$8 pair with $q$9, a probability distribution, and supporting text spans.
• Student Checker: Student LLMs are trained using three-way cross-entropy loss to approximate teacher verdicts:

$D = \{d_1, \ldots, d_k\}$0

• Multiple student models ($D = \{d_1, \ldots, d_k\}$1) are ensembled with F1-weighted class-specific reliability for minority classes (e.g., contradiction).
• Score for class $D = \{d_1, \ldots, d_k\}$2:

  $D = \{d_1, \ldots, d_k\}$3

  where $D = \{d_1, \ldots, d_k\}$4.

3.2 Knowledge Graph Consistency

Claims are aligned to DRKG triples through surface-form and relation matching:

• String Alignment: $D = \{d_1, \ldots, d_k\}$5 computes normalized similarity between claim and DRKG triple, maximizing across all candidates for $D = \{d_1, \ldots, d_k\}$6.
• Embedding Plausibility: TransE-based scoring for $D = \{d_1, \ldots, d_k\}$7 triples:

$D = \{d_1, \ldots, d_k\}$8

Aggregated for $D = \{d_1, \ldots, d_k\}$9 as $k$0.

• Fused KG Support: The overall KG signal for $k$1:

$k$2

with $k$3.

4. Signal Aggregation and Diagnostics

The fusion of NLI and KG signals enables claim-level support quantification:

• Calibrated Support Score: $k$4 is the fused support; $k$5 binarizes support.
• Answer-Level Metrics: For $k$6 claims in answer $k$7,
  • Faithfulness: $k$8
  • Hallucination: $k$9
  • Safety-critical error (for safety-critical claims): $a$0
  • Retrieval diagnosis and context precision: based on claim–evidence linkage and gold reference coverage.

This multi-signal aggregation allows systematic separation of generation-induced versus retrieval-induced errors, facilitating forensic analysis of RAG model weaknesses.

5. Model Distillation and Inference Efficiency

All components are distilled into efficient biomedical LLMs for scalable deployment:

• Claim Extractor and Checker: Compact extractors (Meditron3-8B, Med42-Llama3-8B) are trained via standard SFT/LoRA (e.g., $a$1, $a$2).
• Inference Latency: Pipeline eliminates expensive teacher (GPT-4) calls at inference, running end-to-end in a few seconds per question–answer pair with comparable accuracy.
• Ensemble Reliability: F1-weighted ensembling of multiple checkers boosts detection of rare contradictions and improves macro-F1, which reaches ≈ 60.5% (Table 1 in (Ji et al., 10 Jan 2026)).

6. Empirical Evaluation and Benchmarks

MedRAGChecker has been evaluated across four biomedical QA datasets:

Generator	Faith (%)	Halluc (%)	SafetyErr (%)
Med42-Llama3-8B	85.3	6.3	6.8
PMC-LLaMA-13B	55.1	31.4	34.2

• Datasets: PubMedQA, MedQuAD, TREC LiveQA, MedRedQA.
• Retrieval: BM25/dense PubMed-based RAG, $a$3–10.
• Metrics: Claim-level NLI acc/Macro-F1, answer-level Faith, Halluc, SafetyErr, ClaimF1, CtxPrec.
• Key Findings: KG consistency fusion reduces hallucination and safety error rates on safety-critical claims by 10–20% relative to NLI-only baselines. Human rating alignment improves with KG-induced flips (accuracy rise from 63.4% to 69.8%).

These results highlight the differentiated risk profiles among contemporary biomedical RAG models and validate MedRAGChecker’s fine-grained error localization (Ji et al., 10 Jan 2026).

7. Limitations and Prospects

MedRAGChecker’s major limitations include:

• Teacher/Extraction Bias: GPT-4.1 supervision exhibits domain-dependent noise; downstream models may inherit extraction granularity or bias.
• Contradiction Sensitivity: Detection remains a minority-class challenge; subtle logical contradiction errors may be missed.
• Knowledge Graph Coverage: DRKG does not comprehensively cover biomedical relations; linking may fail on paraphrased or multihop statements.
• Hyperparameter Sensitivity: Fusion and support thresholds are tuned on single development sets, potentially requiring alteration for new sub-domains.

Ongoing directions entail strengthening KG linkage (ontology normalization, multihop reasoning), development of targeted contradiction detectors, joint claim extraction and verification models, and the addition of passage-level explainers for interactive, human-in-the-loop verification (Ji et al., 10 Jan 2026).

---

References

• "MedRAGChecker: Claim-Level Verification for Biomedical Retrieval-Augmented Generation" (Ji et al., 10 Jan 2026)