AI evidence extraction

Evidence Graph Consistency in Retrieval-Augmented Generation: A Model-Dependent Analysis of Hallucination Detection

Jianru Shen · arXiv (Cornell University) · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Computational experimental design with structured graph-based feature extraction on RAG responses.

Main result

The study reveals a critical model-dependent split in hallucination detection effectiveness: "EGC shows the expected diagnostic direction for Llama-2 models but exhibits systematic reversal for GPT-4, GPT-3.5, and Mistral-7B." Specifically, for Llama models all three show ∆m > 0 (grounded answers score higher), while GPT-4 shows the strongest reversal at ∆m = -0.038. When direction-corrected, AUROC improves from 0.556 to 0.669, demonstrating that "EGC carries direction-heterogeneous diagnostic information across model families rather than a uniformly weak signal."

Research paradigm

Empirical/Computational (positivist with model-dependent analysis framework)

Author conclusions

The authors conclude: "The model-family split revealed by EGC carries a broader implication: hallucination is not a monolithic phenomenon. The evaluated Llama-2 models tend to hallucinate by generating claims that are structurally disconnected from the evidence, making the gap detectable through graph topology. The evaluated GPT-class models hallucinate differently, producing fluent, evidence-proximate text that defeats embedding-based structural checks. This distinction suggests that hallucination detection cannot rely on a single signal across all deployment contexts. Effective systems will likely need to match their detection strategy to the characteristics of the underlying model." Furthermore: "EGC should not be used as a universal stand-alone hallucination detector. Instead, it is best interpreted as a lightweight structural diagnostic layer whose reliability depends on the hallucination style of the underlying model."

Risk of bias

Class imbalance (1:2.4 ratio, 29.6% hallucinated) addressed via balanced class weights but may still affect model generalization; Fixed similarity threshold τ=0.4 not optimized per domain or embedding model; Single encoder (all-MiniLM-L6-v2) may introduce embedding-model-specific artifacts; Sentence-level claim segmentation may conflate supported and unsupported claims within single sentences; Binary response-level labeling discards span-level granularity that could reveal more nuanced patterns; Single domain bias (open-domain QA only, no summarisation or data-to-text tasks tested); Embedding model bias (all-MiniLM-L6-v2 used exclusively; generalization to other encoders untested); Fixed similarity threshold (τ=0.4 not learned, potential domain-specific mismatch); Sentence-level claim segmentation (coarse approximation conflating supported/unsupported claims within sentences); Class imbalance addressed via balanced logistic regression but original 1:2.4 ratio preserved in evaluation; Binary response-level labeling discarding span-level granularity from original annotations; Selection bias: Single dataset from single domain (open-domain QA only); generalisation to summarisation and data-to-text tasks not established; Information loss bias: Binary response-level labeling discards span-level annotation detail including hallucination span length, type, and count within responses; Threshold arbitrariness: Fixed threshold τ = 0.4 not learned; choice may not generalise to other domains or embedding models; Encoder bias: All node encodings use single embedding model (all-MiniLM-L6-v2); different encoders may yield different similarity distributions; Segmentation bias: Sentence-level claim segmentation is coarse approximation; may conflate supported and unsupported claims within single sentence; Class imbalance handling: Addressed via balanced class weights rather than domain-specific examination of whether equal weight is appropriate

Limitations

The authors state: "although Section V-E shows that τ = 0.4 achieves perfect directional consistency across all six models, the threshold is not learned and may not generalise to other domains or embedding models
Second, claim segmentation by sentence boundary is a coarse approximation
atomic claim extraction via natural language inference would provide finer-grained claim nodes
Third, cross-evidence agreement provides limited aggregate signal, although its strong reversed association for GPT-4 suggests that E-E structure may capture model-specific behaviour not reflected in pooled evaluation
Finally, all results are reported on a single dataset from a single domain (open-domain QA)
generalisation to other RAG tasks such as summarisation or data-to-text generation is not established." Additionally, "our evaluation converts RAGTruth's span-level annotations into a binary response-level label

Open questions raised

1. Model-adaptive thresholds: Current fixed τ=0.4 may not generalize to other domains or embedding models; learning τ as a function of model family characteristics is needed. 2. Finer-grained claim segmentation: Atomic claim extraction via NLI would improve over sentence-level segmentation. 3. Cross-task generalization: Results limited to QA; testing on RAGTruth summarization and data-to-text splits needed. 4. Combining structural and semantic signals: Hybrid detectors combining EGC features with NLI-based entailment, token-level confidence, or internal model states could address GPT-class failure modes.
Model-adaptive thresholds: Current fixed τ=0.4 not learned; model-adaptive or learned threshold accounting for embedding similarity distribution needed
Finer-grained claim segmentation: Atomic claim extraction via NLI (as in FActScore) needed instead of sentence-level boundaries
Cross-task generalisation: Need to test model-family split on RAGTruth summarisation and data-to-text splits beyond QA
Combining structural and semantic signals: Hybrid approach integrating internal model states, token-level confidence, or NLI-based entailment for stronger models
Generalisation to other embedding models: τ=0.4 and findings may not transfer beyond all-MiniLM-L6-v2

Data: RAGTruth corpus (5,767 responses from six LLMs on 989 unique questions from MS MARCO). Official train/test splits provided: 4,892 train, 875 test. Includes word-level hallucination annotations categorized into: evident conflict, subtle conflict, evident baseless information, and subtle baseless information.; RAGTruth: Available via official release; 17,790 naturally generated responses from six LLMs across QA, data-to-text, and summarisation tasks with word-level hallucination annotations. QA subset used: 5,767 responses (4,892 train, 875 test) from 989 unique questions from MS MARCO.; RAGTruth (question answering subset): 5,767 responses from six LLMs, 989 unique questions from MS MARCO, official train/test splits (4,892 train, 875 test). Paper uses released responses without new generation.Extracted from: pdfAgreement 60%

Explore related topics

Hallucination Control Research Integrity Empirical Study