TL;DR
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.
A reproducible estimate of diagnostic retrieval accuracy across four widely used model configurations is provided to establish a baseline for further clinical validation and establish a baseline for further clinical validation.
Large Language Models (LLMs) such as ChatGPT and Gemini are increasingly used to answer medical and psychological questions, yet systematic evaluations across domains with differing reasoning demands remain limited. We present a multi-domain expert evaluation of two state-of-the-art models, ChatGPT Pro (v5.2) and Gemini 3 Pro, across three healthcare domains: Gynecology, Pathology, and Psychology. We curated 300 open-ended, realistic questions, 100 per domain, designed to elicit clinical reasoning, mechanistic interpretation, and conceptual explanation. Responses were independently scored by domain experts using a standardized five-point rubric. Results reveal domain-dependent performance patterns, with both models performing well in guideline-aligned advisory tasks in Gynecology, lower in mechanistic diagnostic contexts in Pathology, and showing divergent strengths in conceptual psychology questions. Quantitative and qualitative analyses highlight recurring limitations in contextual nuance, mechanistic depth, and safety framing. These findings underscore the importance of domainstratified evaluation and expert oversight when deploying LLMs for healthcare information and provide a reproducible framework for future assessments.
Bottom-up incremental scoring showed the closest alignment with human assessment in clinical AI evaluation, underscoring the need for standardised prompt architectures in clinical AI evaluation.
AI-generated answers in high-stakes domains are often fluent but difficult to verify, especially when they contain multi-step reasoning rather than a single final answer. We propose a reasoning-based, reference-free framework for auditing LLM-generated outputs. The method decomposes a generated reasoning trace into segments, labels local premise-target relations using Natural Language Inference (NLI), and organizes these relations into a hypergraph. A deterministic backward AND-OR search then assigns segment-level audit labels that indicate how each segment is grounded within the generated response. We evaluate the framework in two settings: deductive mathematical reasoning with Hard2Verify, and open-ended medical reasoning with UroReason, a new physician-annotated benchmark of LLM reasoning traces from real clinical cases. Across these settings, our NLI-hypergraph audit provides a more reliable reference-free evaluation signal than direct LLM-as-judge baselines. In the clinical setting, state-of-the-art LLM judges often fail to identify problematic reasoning segments, over-accepting fluent but weakly grounded responses. Our results show that QA evaluation should account for how inferential relations compose across a reasoning trace, rather than relying only on final answers or LLMs as verifiers. UroReason will be made available through an API, and our code will be released as open source.
Background: The reliable integration of large language models (LLMs) into neuroimaging data extraction workflows remains unresolved. Prior benchmarking shows that exact-match accuracy underestimates LLM extraction performance, but whether inter-model consensus and variable complexity can guide automation remains unclear. We evaluated whether inter-model consensus can serve as a confidence signal for human–artificial intelligence (AI) extraction and can guide complexity-stratified workflow triage. Methods: Four frontier LLMs were queried via OpenRouter with an identical zero-shot structured prompt to extract 22 predefined variables from 91 peer-reviewed neuroimaging AI articles, yielding 2002 article–variable items per model. Variables were stratified a priori into low- (n = 7), medium- (n = 8), and high-complexity (n = 7). Performance was compared with an expert reference using exact-match and semantic-equivalence accuracy. Item-level consensus and five triage strategies characterized the efficiency–accuracy trade-off. Results: Semantic-equivalence accuracy converged to 80.5–83.4% across models despite approximately ten percentage-point exact-match differences. Unanimous 4/4 consensus occurred in 45.6% (910/1994) of items, with exact-match accuracy of 85.8%, rising to 95.3% after semantic normalization; however, 14.2% still failed to match the reference. Reliability was complexity-dependent: 96.6% for low-complexity variables, 73.2% for medium-complexity variables, and 38.1% for high-complexity variables. A hybrid strategy auto-accepting 4/4 items and routing 3/4 items to rapid verification reduced estimated review effort by approximately 59%. Conclusions: Inter-model consensus is useful, but it is incomplete and depends on variable complexity. We show that LLM-assisted extraction in neuroimaging AI is a complexity-stratified workflow design problem: low-complexity neuroimaging variables may be selectively automated, while medium-complexity variables require rapid verification, and high-complexity methodological variables should remain human-led.
Specialized thoracic-surgery questions require the integration of multi-factor clinical relationships within text, yet general-purpose large language models (LLMs) may underperform on such exam-style benchmarks. We constructed a DK-LLM agent by embedding curated medical textbook knowledge into a LangChain-based framework to support domain-specific reasoning. The model was tested on a 56-item thoracic-surgery examination question set in a restricted text-only setting without internet browsing or external tools and was compared with generic LLM baselines, three thoracic surgeons, and three non-expert engineers. Examination score and error patterns were assessed. The knowledge-augmented DK-LLM configuration showed an 11.8-point examination-score advantage over the version without the local knowledge base. Commercial LLM-based agents outperformed the open-source baselines and non-expert participants on this question set, whereas experienced thoracic surgeons achieved the highest scores overall; in the ablation analysis, removing the local knowledge base reduced the examination score by 11.8 percentage points. Embedding domain-specific knowledge into LLMs may improve performance on specialized exam-style thoracic-surgery questions on this text-only benchmark. However, the present 56-item evaluation does not establish clinical equivalence, diagnostic accuracy in practice, multimodal competence, or readiness for real-world clinical decision support.