How Often Do Large Language Models Agree with Each Other—And with the Truth? A Consensus- and Complexity-Stratified Analysis of Data Extraction for Neuroimaging AI

Multi-Axial Analysis of Clinical Reasoning in Large Language Models: Inter-Verifier Disagreement and Its Implications for Automated Evaluation

A single LLM verifier lacks sufficient reliability to serve as a stand-alone judge of clinical reasoning at scale, and that structured human oversight remains essential.

Automated Interpretation of EEG Reports Using a Large Language Model with Structured Confidence Outputs

Large language models generate diagnostic likelihood ratios with low mean bias but wide dispersion.

Accurate, context-appropriate likelihood ratios (LRs) are needed for Bayesian diagnosis, but empirical LRs are sparse because diagnostic accuracy studies are costly and context-dependent. We evaluated whether large language models (LLMs) can generate LR outputs that align with literature-reported values. We compared LR outputs generated by three OpenAI models (GPT-4o, o3, and GPT-5) with all literature-reported values curated in TheNNT.com dataset. A few-shot prompt was used to elicit numerical LR outputs, and agreement was assessed using Bland-Altman analyses to evaluate mean bias and multiplicative limits of agreement. A total of 700 reported LRs across 30 conditions were compiled, most involving signs or symptoms (59%), historical elements (19%), or test results (16%). All models demonstrated negligible mean bias. GPT-5 had the narrowest 95% limits of agreement (0.26×-3.70×) compared with o3 and GPT-4o. These findings indicate that LLM-generated LRs can be centered near literature values on average, but they are not interchangeable with published estimates for individual high-stakes decisions. Clinical use, including use for unstudied findings, would require prospective, context-specific diagnostic-accuracy validation.

Performance of large language models in data extraction for evidence synthesis: A systematic review.

OBJECTIVE Data extraction is among the most resource-intensive and error-prone stages of systematic review production. Large language models (LLMs) offer potential for automating or semi-automating this process, yet their performance characteristics remain incompletely characterised. This systematic review aimed to comprehensively evaluate LLM accuracy, reliability, and efficiency for data extraction in evidence synthesis, and to identify optimal implementation strategies. METHODS We searched PubMed, Embase, Web of Science, and preprint servers (medRxiv, arXiv) through December 2025. Studies were eligible if they evaluated one or more LLMs for data extraction against a human reference standard and reported quantitative performance metrics. Two reviewers independently extracted data and assessed methodological quality using PROBAST + AI and reporting completeness using TRIPOD-LLM. Narrative synthesis was performed due to substantial heterogeneity precluding meta-analysis. RESULTS Twenty-seven studies met inclusion criteria, evaluating models including GPT-4/4o (n = 15), Claude versions 2-3.5 (n = 10), Gemini (n = 3), and open-source alternatives including Llama, Mistral, Qwen, and DeepSeek. Overall accuracy ranged from 47% to 99.9%, with substantial heterogeneity by task type and data granularity. Categorical and string variables were extracted more reliably (74-96%) than numerical data (47-88%). Claude 3.5 Sonnet achieved high accuracy in an assistive workflow (91.0%; 95% CI: 90.4-91.6%), exceeding human-only extraction (89.0%). Claude models outperformed GPT in head-to-head comparisons (OR 1.70 for event counts). Omissions were the dominant error type (60-74%), with hallucination rates of only 0.08-6%, challenging widespread fabrication concerns. Time savings of 33% to 87% were reported, although most included studies did not quantitatively assess efficiency. Methodological quality was generally robust, with 74.1% of studies rated low risk of bias under PROBAST + AI. Mean TRIPOD-LLM compliance was 88.5%, though gaps in inference settings and model version documentation were common. CONCLUSION LLMs demonstrate promising but variable performance for data extraction in evidence synthesis. Current evidence supports their integration as assistive tools within dual-extraction workflows requiring human verification, rather than as autonomous extractors. Categorical data is extracted more reliably than numerical outcomes, and few-shot prompting with structured output formats consistently improves performance. Standardised benchmarks and prospective comparative studies remain priorities for future research.

Confident but Unreliable: A Behavioral Safety Audit of Vision-Language Models on Brain MRI

Vision-language models (VLMs), including medical specialists, are increasingly proposed for medical imaging, yet their stated confidence is rarely evaluated separately from correctness. We use brain MRI as a controlled, high-stakes testbed for a broader failure mode in frontier multimodal systems: models can appear competent while lacking reliable self-knowledge. We present an automatically graded behavioral audit and pilot study of six instruction-tuned VLMs (five general-purpose and one medical specialist) on 4,102 images (4,032 axial/coronal/sagittal MRI slices from 250 subjects plus 70 non-brain/noise controls), with labels derived from public metadata and released expert segmentation masks rather than new human annotation. Across models, answer coverage is near-complete, but verbalized-confidence calibration is poor: ECE ranges from 0.27 to 0.40, mean confidence on incorrect answers ranges from 0.82 to 0.97, and 33-46% of answered items are high-confidence errors. The most accurate model is also the most confident on its errors, while a base/specialist family contrast suggests that medical adaptation improves tumor-presence detection without improving confidence reliability. Open-ended diagnostics further show that hallucination and abstention vary separately from multiple-choice accuracy. These findings argue that medical-image VLM evaluation should report verbalized-confidence reliability, confident error, hallucination, and abstention alongside accuracy.

The Computational Basis of Confidence in Large Language Models

Reliable confidence -- the probability that a model's own answer is correct -- is essential for the trustworthy deployment of language models. Existing work has largely evaluated confidence by how well it predicts correctness and whether it is calibrated, leaving open a more fundamental question: what does the confidence signal itself represent? Answer logits may reflect a latent decision variable sufficient to compute normative confidence, or instead a heuristic preference signal that combines the available evidence in a non-Bayesian manner. We address this using statistical decision confidence (SDC), a normative framework from computational neuroscience. Treating the answer-logit difference (LD) as a candidate readout of the latent decision variable, we test the qualitative signatures predicted by SDC. Across three perceptual discrimination tasks and a memory-based decision task, spanning three multimodal non-reasoning models and one reasoning model, LD satisfied these signatures -- including the diagnostic correct/error folded-X pattern -- showing that, in these settings, answer logits behave as monotonic readouts of a latent decision variable rather than heuristic preference scores. In complex visual reasoning, LD continued to predict correctness beyond objective task difficulty, but the full geometric signatures of SDC were absent, illustrating the current boundary of the framework when explicit normative process models are unavailable. These results provide a computational account of confidence in multimodal language models, delineate when answer logits behave as readouts of a latent decision variable, and establish SDC as a unifying framework for studying confidence across biological and artificial intelligence.