It also achieved 91.6% factual accuracy for clinical-entity recognition. These experimental results support the technical value of structured medical knowledge, but they do not establish improved patient outcomes or readiness for unsupervised clinical use. HalluciGuard-Eye was built on the LLaVA-1.5 architecture, using a ViT-L/14 vision encoder fine-tuned on fundus and OCT data and a Vicuna-13B language model.

Its dual-stream design separates two tasks that conventional multimodal language models may handle imperfectly: determining whether a statement corresponds to what is visible in an image and determining whether it agrees with established ophthalmic knowledge. The Visual-Semantic Alignment Module compares visual tokens with claim-level text through bidirectional attention and contrastive learning. Its visual representation retains patch-level details, quadrant-level context and a global token, allowing the system to evaluate small lesions, local vascular structures and the overall distribution of disease.

In an ablation study, removing this module increased the EyePACS hallucination rate from 22.2% to 26.6% and lowered the Hierarchical Factuality Score from 81.7 to 76.4. The largest gain from visual-semantic alignment occurred for small lesions: lesion F1 rose by 14.9 percentage points among lesions occupying less than 0.25% of an image, compared with a 5.6-point gain for lesions occupying at least 1%. The Clinical Factuality Grounding Network supplies the system with structured evidence about anatomical structures, pathology, diseases, symptoms, procedures and imaging findings.

The underlying Ophthalmic Clinical Knowledge Graph contains 47,832 entities and 156,294 semantic triplets, organized into hierarchical, causal, spatial and clinical relationships. Its sources include guidance from the American Academy of Ophthalmology, the European Society of Retina Specialists and the International Council of Ophthalmology, the ICD-11 classification, standardized diabetic-retinopathy and age-related eye-disease grading systems, and ophthalmology literature indexed in PubMed. Conflicting guidance is prioritized by the most recent applicable guideline, while unresolved conflicts are marked for senior-expert review rather than silently combined.

An independent stratified audit of 3,000 graph triplets found that 92.7% were supported, with a 95% confidence interval of 91.7% to 93.6%. The authors note that this estimate describes the audited sample rather than every edge in the graph. For each candidate response, HalluciGuard-Eye recognizes clinical entities, assigns greater retrieval weight to claims involving vision-threatening lesions or referral relationships, and searches relation-aware portions of the graph.

It then expands the retrieved subgraphs through multi-hop traversal before reranking candidate answers. The selected response is generated with a gated mechanism that combines language-model states with retrieved evidence during decoding. The researchers used a two-hop retrieval depth after a sensitivity analysis found that it offered the best balance between critical-evidence recall and hallucination rate.

Increasing the depth to four hops raised recall slightly, from 87.6% to 89.8%, but reduced the factuality score and increased latency. The graph was central to performance: removing it increased the EyePACS hallucination rate to 29.8% and lowered HFS to 72.3, a reduction of 9.4 points from the full system. Removing the Clinical Factuality Grounding Network produced an intermediate decline, with a 28.1% hallucination rate and an HFS of 74.8.

The system was evaluated on four public ophthalmic imaging datasets. EyePACS contained 35,126 labeled fundus photographs, including a 3,512-image test partition and a separately annotated subset of 1,500 images used for lesion-level assessment. Messidor-2 supplied 1,748 external fundus images without model fine-tuning, while OCTDL contained 2,064 OCT B-scans and the Kermany OCT dataset contained 84,484 images.

Comparisons included LLaVA-Med, Med-Flamingo, HuatuoGPT-Vision and Qwen-VL-Chat, as well as retrieval- and hallucination-focused systems including Self-RAG, V-RAG, MedRAG and MediHallDetector. On Kermany OCT, HalluciGuard-Eye recorded a 23.8% hallucination rate versus 41.2% for LLaVA-Med, again corresponding to a 42.2% relative reduction. For diabetic-retinopathy grading on EyePACS, it achieved 89.6% sensitivity and 92.1% specificity, exceeding MedRAG by 2.4 and 2.7 percentage points, respectively.

Its AUC was 0.951 and quadratic weighted kappa was 0.873. Five-seed testing produced low dispersion, with a hallucination rate of 22.2% plus or minus 0.5% and sensitivity of 89.6% plus or minus 0.4%. Retrieval tests indicated that the system’s gains came from combining the graph with claims about what mattered clinically.

Replacing an unweighted graph query with a salience-weighted, relation-aware query increased critical-evidence recall at 20 from 78.1% to 87.6% and evidence normalized discounted cumulative gain at 20 from 0.742 to 0.811. At the same time, the hallucination rate fell by 2.7 percentage points and HFS rose by 2.5 points. Recall for vision-threatening lesion claims rose from 74.3% to 88.1%.

Masking progressively larger portions of the graph produced a graded deterioration: when only 50% of critical relations were retained, critical-evidence recall declined to 71.0%, the hallucination rate rose to 28.9% and HFS fell to 75.4. This sensitivity analysis supports the graph’s contribution, but it also shows that HalluciGuard-Eye depends on the coverage and quality of its curated medical knowledge. Open-ended reports were assessed on 1,500 expert-annotated cases rather than reduced only to a disease label.

HalluciGuard-Eye reached 90.8% lesion-presence F1, 89.2% lesion-location accuracy, 86.7% severity accuracy, 92.6% negation F1, 86.1% hedging appropriateness and 91.4% clinical-entity F1. Its largest improvement over MedRAG was in severity accuracy, at 8.5 percentage points. The study’s Hierarchical Factuality Score separates reliability into anatomical accuracy, pathological characterization and validity of clinical recommendations.

HalluciGuard-Eye scored 86.4, 79.0 and 81.4 on those three components in EyePACS, producing a composite score of 81.7. Pathological characterization remained the weakest component for both the proposed system and its comparators. The ranking did not change when equal or anatomy-, pathology- or recommendation-heavy weights were applied, although the findings do not establish that the composite score is an externally validated measure of real-world safety.

A retrospective patient-clustered cohort offered a different test because it included both eyes from some participants and examined performance in a clinical collection rather than only a public test split. The analysis covered 512 patients and 928 eyes observed across four calendar years, with 33.8% prevalence of referable diabetic retinopathy. Model, retrieval corpus, prompts and thresholds were reportedly frozen before reference labels were finalized.

HalluciGuard-Eye achieved 88.9% sensitivity and 90.7% specificity for referable disease, with an AUC of 0.946 and a serious-hallucination rate of 8.5%. The corresponding MedRAG values were 85.0% sensitivity, 88.1% specificity, an AUC of 0.920 and a 14.4% serious-hallucination rate. The analysis accounted for correlation between fellow eyes through patient-clustered bootstrapping and generalized estimating equations.

Sensitivity was highest for high-quality images at 91.7%, compared with 88.9% for moderate-quality and 84.1% for low-quality images, illustrating that image quality remained an important influence. Human evaluations also favored the grounded system. Five board-certified ophthalmologists rated 200 responses from each method, assigning HalluciGuard-Eye the highest accuracy and clinical-trust scores.

In a separate crossover experiment, 24 ophthalmologists assessed 480 referral vignettes with and without model assistance. Assisted review increased referral sensitivity from 86.9% to 92.7% and specificity from 88.2% to 91.0%. Serious decision errors declined from 11.5% to 6.9%, while median reading time fell from 52 to 42 seconds.

These are experimental workflow results from a controlled vignette study, not evidence that deploying the system will improve real patient outcomes. The researchers also profiled the engineering burden: for a 256-token report, retrieval added 0.12 to 0.15 seconds, and four-bit inference on a 24-gigabyte Nvidia RTX 4090 produced a latency of 2.47 seconds per report. Overall, the study shows a technically coherent way to reduce fabricated claims and ground image interpretation in explicit clinical relationships, while leaving the prospective, multicenter and patient-outcome testing required for clinical deployment unresolved.

Subject of Research: Benchmark and retrospective evaluation of knowledge-grounded hallucination detection in ophthalmic multimodal AI

Article Title: HalluciGuard-eye: knowledge graph-enhanced retrieval-augmented generation for hallucination detection and mitigation in ophthalmic multimodal large language models

Article References: Wu, G., Ai, C., & Wu, W. (2026). HalluciGuard-eye: knowledge graph-enhanced retrieval-augmented generation for hallucination detection and mitigation in ophthalmic multimodal large language models. Frontiers in Endocrinology, 17, Article 1976190. https://doi.org/10.3389/fendo.2026.1976190

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1976190

Keywords: ophthalmic artificial intelligence, multimodal language models, hallucination detection, diabetic retinopathy, knowledge graphs, retrieval-augmented generation, fundus photography, optical coherence tomography, clinical factuality, retrospective cohort, HalluciGuard-eye, knowledge

The framework, called CLAIM-CAL, was developed by Abhigyan Pal, an independent researcher based in New Delhi, India. Rather than treating a model’s entire response as a single object to be trusted or doubted, CLAIM-CAL decomposes each answer into atomic factual claims, the smallest independently checkable statements an answer contains. A response such as ‘The Eiffel Tower is in Paris and was completed in 1889’ is split into two distinct claims, each of which is then verified on its own merits. This granular approach reflects a core insight of the research: a single answer can be a mixture of accurate and inaccurate statements, and assigning one blanket confidence score to the whole response obscures exactly the information users need most.

The technical pipeline works in five stages. First, an answer model generates an initial response to a question without any external retrieval, deliberately isolating the contribution of self-verification. Second, gpt-4o-mini operating at temperature 0.0 extracts atomic claims using a structured JSON-format prompt, capping extraction at eight claims per answer with a sentence-splitting fallback if extraction fails. Third, each claim is interrogated by three differently framed verification probes: a balanced probe, a support-seeking probe, and a contradiction-seeking probe. Each probe labels the claim as Supported, Contradicted, or Unknown, and the deliberately adversarial framing of the contradiction probe is central to the design, because actively hunting for disconfirming evidence catches confidently wrong statements that a purely supportive check might wave through.

Fourth, the verdicts are converted into risk scores using a transparent hand-set heuristic. Contradiction carries the heaviest penalty because direct evidence against a claim is the strongest signal of unreliability; Unknown claims receive an intermediate penalty because they cannot be confidently verified; and lack of explicit support adds a smaller but nonzero penalty. The claim-level risks are averaged to produce an answer-level risk, which is subtracted from one to yield a raw confidence score. Crucially, the study does not stop there. The fifth stage applies isotonic regression, a non-parametric post-hoc calibration technique that learns a monotonic mapping from raw confidence scores to empirical correctness using a dedicated calibration split of sixty examples. Without this final step, even the sophisticated claim-level signal remains systematically overconfident.

The evaluation used the TruthfulQA generation benchmark, a dataset specifically engineered to expose cases where language models reproduce common human misconceptions. The experiment sampled 200 examples, reserving 60 for calibration and 140 for held-out testing. CLAIM-CAL was benchmarked against four alternatives: direct answering with assumed full confidence, verbal confidence where the model self-reports its certainty, self-consistency based on agreement across five independently sampled answers, and simple self-verification that judges the whole answer at once. On the held-out split, CLAIM-CAL achieved the highest observed accuracy of 0.757 among the evaluated methods, but the more striking finding concerned calibration quality rather than raw accuracy.

Raw CLAIM-CAL, like every other uncalibrated method, remained markedly overconfident, with an Expected Calibration Error of 0.212. ECE measures the bin-weighted gap between what a system claims to know and what it actually gets right. After isotonic calibration, that figure collapsed to 0.038, with a 95 percent bootstrap confidence interval of 0.015 to 0.106, while accuracy held steady. To rule out the possibility that calibration alone explained the advantage, the study then applied the same isotonic procedure to the variable-confidence baselines on the identical calibration split. In this fair head-to-head comparison, the calibrated claim-level method retained the strongest point estimates for both ECE and Brier Score, although the author is careful to note that paired bootstrap intervals overlapping zero mean the ECE advantage over the closest calibrated baselines should be interpreted cautiously rather than as statistically decisive superiority.

Where the approach truly shines is in selective answering, the practical scenario in which a deployed system must decide when to answer and when to defer. At a fixed 0.7 confidence threshold, CLAIM-CAL plus calibration achieved selective accuracy of 0.824 while retaining 0.893 coverage, meaning the system answered nearly 90 percent of questions while being right more than 82 percent of the time on the ones it chose to answer. Self-consistency, by contrast, reached comparable selective accuracy only at a coverage of 0.207, essentially answering so few questions that its apparent reliability becomes operationally useless. After calibration, self-consistency’s confidence values never exceeded 0.667, leaving it with zero coverage at the threshold entirely. This coverage-versus-accuracy trade-off is a point the study emphasizes repeatedly: a method can look trustworthy simply by refusing to engage.

The paper is notable as much for its methodological candor as for its results. Because the same model family handled answer generation, claim verification, and correctness judging, correlated errors could inflate apparent performance, a confound the author explicitly flags. A second-pass robustness check using an independent stricter judge prompt on 50 sampled evaluations achieved 96 percent agreement and a Cohen’s kappa of 0.896, which is strong consistency but not human validation. An ablation study confirmed that removing contradiction probing produced the worst calibration error among the variants, underscoring the value of adversarial verification, while removing claim decomposition lowered accuracy and weakened the reliability signal. The error analysis also showed the method’s honest limits: twenty-two incorrect answers still slipped through above the 0.7 confidence threshold, proving that CLAIM-CAL improves calibration and risk awareness without guaranteeing correctness.

The implications reach well beyond a single benchmark. The framework is designed to be complementary to prompt engineering and retrieval-augmented generation: better prompts can reduce initial errors, retrieved evidence can ground claims externally, and CLAIM-CAL can sit on top of either, estimating whether the final answer deserves user trust. A natural next step is verifying extracted claims directly against retrieved passages, converting the self-verification layer into a retrieval-grounded reliability check. Future work identified in the study includes scaling beyond 200 examples, testing multiple model families for generator, verifier, and judge roles, learning the risk weights from data rather than fixing them by hand, and integrating human review for low-confidence answers. The deeper conclusion is a reframing of what reliability means for artificial intelligence: a system should not be judged only by whether its answers are correct, but by whether its expressed confidence honestly reflects the probability of correctness. In an era when fluent language is too easily mistaken for truth, teaching models to know what they do not know may prove as important as teaching them what they do.

Subject of Research: Claim-level self-verification and uncertainty calibration for improving the reliability of large language models

Article Title: Improving reliability of large language models via claim-level self-verification and uncertainty calibration

Article References: Pal, A. (2026). Improving reliability of large language models via claim-level self-verification and uncertainty calibration. Discover Artificial Intelligence, 6(1), Article 1132. https://doi.org/10.1007/s44163-026-02240-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02240-w

Keywords: large language models, uncertainty calibration, self-verification, hallucination detection, TruthfulQA, isotonic regression, selective answering, Expected Calibration Error, reliability estimation, natural language processing, AI trustworthiness, confidence estimation