Benchmarking Multimodal Large Language Models for Forensic Science and Medicine: A Comprehensive Dataset and Evaluation Framework
Bibliographic record
Abstract
Abstract Background Multimodal large language models (MLLMs) have demonstrated substantial progress in medical and legal domains in recent years; however, their capabilities from the lens of forensic science—a field that is at the intersection of complex medical reasoning and legal interpretation, with conclusions critiqued by judicial scrutiny—remains largely unexplored. Forensic medicine uniquely depends on the accurate integration of often ambiguous text and visual information, yet systematic evaluations of MLLMs in this setting are lacking. Methods We conducted a comprehensive benchmarking study of eleven state-of-the-art MLLMs, including proprietary (GPT-4o, Claude 4 Sonnet, Gemini 2.5 Flash) and open-source (Llama 4, Qwen 2.5-VL) models. Models were evaluated on 847 examination-style forensic questions drawn from various academic literature, case studies, and clinical assessments, covering nine forensic subdomains. Both text-only and image-based questions were included. Model performance was assessed using direct and chain-of-thought prompting, with automated scoring verified through manual revision. Results Performance improved consistently with newer model generations. Chain-of-thought prompting improved accuracy on text-based and choice-based tasks for most models, though this trend did not hold for image-based and open-ended questions. Visual reasoning and complex inference tasks revealed persistent limitations, with models underperforming in image interpretation and nuanced forensic scenarios. Model performance remained stable across forensic subdomains, suggesting topic type alone did not drive variability. Conclusions MLLMs show emerging potential for forensic education and structured assessments, particularly for reinforcing factual knowledge. However, their limitations in visual reasoning, open-ended interpretation, and forensic judgment preclude independent application in live forensic practice. Future efforts should prioritize the development of multimodal forensic datasets, domain-targeted fine-tuning, and task-aware prompting to improve reliability and generalizability. These findings provide the first systematic baseline for MLLM performance in forensic science and inform pathways for their cautious integration into medico-legal workflows.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.020 | 0.040 |
| Meta-epidemiology (narrow) | 0.003 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.003 |
| Bibliometrics | 0.006 | 0.004 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.003 | 0.003 |
| Open science | 0.005 | 0.005 |
| Research integrity | 0.004 | 0.003 |
| Insufficient payload (model declined to judge) | 0.004 | 0.003 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".