A recent study evaluating multimodal Large Language Models for forensic image analysis has revealed that while these systems can successfully triage large datasets, their reliance on surface-level visual cues like uniforms often leads to high rates of false-positive identification. By testing Google’s Gemini 1.5 Pro, the open-source LLAVA model, and OpenAI’s GPT-4o on a dataset of 434 mobile device images, researchers found that these models achieved high recall—up to 0.99—but struggled with moderate precision, largely due to an inability to distinguish real military personnel from highly realistic mannequins. The findings, published in the International Journal of Networked and Distributed Computing, indicate that baseline multimodal models frequently overemphasize uniforms and equipment, demonstrating a need for more rigorous verification methods in AI-assisted forensic investigations.
To address these shortcomings, the research team introduced and tested two agentic orchestration frameworks: CrewAI, which utilizes collaborative sub-agents, and LangGraph, which employs a deterministic directed acyclic graph pipeline. These frameworks were designed to decompose the forensic pipeline into specialized sub-tasks including provenance validation, perception, mannequin discrimination, and evidence-grounded attribution. By embedding the models within these structured reasoning workflows, the researchers aimed to transform free-form inference into a controlled, auditable process. The results showed that these agentic architectures significantly outperformed baseline single-pass inference, with LangGraph reaching a precision of 0.90 and reducing mannequin-related false positives to 8 percent, compared to the average baseline false positive rate of 27 percent.
The study design involved a carefully curated dataset composed of 198 images of military personnel, 99 images of mannequins dressed in military gear, and 137 non-military images acting as a negative control group. To simulate real-world forensic conditions, all images were transferred via WhatsApp and extracted using iMazing software, mimicking the recovery processes common in mobile device investigations. Each image underwent a multi-stage analysis where the LLMs were prompted to provide comprehensive descriptions, which were then converted into a structured JSON format to evaluate metrics including precision, recall, and accuracy. This systematic approach allowed the team to measure not just detection capabilities, but also the models’ proficiency in identifying contextual attributes such as country of origin and military unit names.
Quantitative comparisons revealed distinct performance characteristics among the base models. Gemini 1.5 Pro achieved an accuracy of 0.793 in initial military personnel identification, matching the performance of LLAVA, while GPT-4o followed with 0.770 accuracy. However, Gemini 1.5 Pro showed superior contextual understanding, attaining an accuracy of 0.787 for country identification, whereas GPT-4o and LLAVA trailed significantly with scores of 0.385 and 0.121, respectively. Despite these differences, all baseline models exhibited a shared vulnerability to mannequin misclassification, with each model misclassifying over 85 percent of the mannequin-containing images, confirming that current multimodal models prioritize visual uniform patterns over human authenticity markers like skin texture or body language.
The introduction of agentic orchestration frameworks fundamentally altered these outcomes by enforcing evidence-based reasoning and enabling conservative abstention logic. The CrewAI configuration, modeled on collaborative agent interaction, achieved an overall detection precision of 0.88, while the LangGraph approach provided a more stable, deterministic output. These frameworks utilized dedicated agents for tasks like mannequin discrimination, which proved highly effective at filtering false positives. While the integration of these layers led to a slight reduction in total recall, the researchers noted that this was a deliberate consequence of the system’s new ability to abstain from making conclusions when confidence levels were low, rather than a degradation in detection quality.
Qualitative analysis underscored the practical advantages of agentic systems in forensic environments. While standard LLMs often defaulted to vague, generic descriptions when faced with complex visual conditions—such as poor lighting, non-standard poses, or occlusions—the agentic pipelines generated reports accompanied by explicit evidence citations and clear decision traces. CrewAI’s narrative-style reasoning provided contextual depth suitable for exploratory analysis, while LangGraph’s structured output formats were specifically optimized for audit-ready documentation. These features offer a path toward meeting forensic standards, as the outputs provide reproducible steps that allow human analysts to verify the model’s reasoning rather than relying on opaque, probabilistic classifications.
Despite the successes of these agentic architectures, the study identified persistent challenges that remain for future research. Additionally, the increase in processing latency associated with agentic pipelines, which was approximately 2.8 times higher than single-pass inference in the LangGraph configuration, represents a computational trade-off that must be balanced in large-scale investigations. However, the researchers emphasize that this increase in latency is offset by the gain in forensic trustworthiness and the reduction of speculative hallucinations.
In conclusion, the study demonstrates that the application of multimodal LLMs in digital forensics should move beyond simple, reactive inference. By shifting toward proactive, structured forensic reasoning through agentic orchestration, investigations can achieve higher analytic precision and improved evidentiary defensibility. The ability to transform raw image data into structured, evidence-backed reports marks a significant step toward the reliable deployment of AI in national security and criminal investigations.
To investigate the viability of automated image assessment, the researchers established a rigorous experimental framework centered on 434 mobile device images collected from an iOS file system backup. These images were originally transferred via WhatsApp, a common channel in digital investigations, and extracted using iMazing software to preserve authentic data characteristics. The dataset was segmented into three distinct groups: 198 images of actual military personnel, 99 images of military-themed mannequins, and 137 images of civilians as a negative control. By including realistic mannequins, the study directly challenged the models’ ability to distinguish between genuine human figures and inanimate objects, a persistent hurdle in computer-vision forensics that complicates field operations.
The evaluation incorporated both individual multimodal Large Language Models and two structured agentic orchestration architectures. For the single-pass baseline, the researchers evaluated Google’s Gemini 1.5 Pro, the open-source LLAVA model, and OpenAI’s GPT-4o. The team utilized Python and the Langchain library to query these models, requesting comprehensive scene descriptions that included identifiable individuals, uniforms, equipment, and military context. These descriptions were subsequently converted into a structured JSON format using OpenAI’s o3-mini-2025-01-31 model to facilitate precise metric measurement. This secondary processing step allowed the researchers to extract uniform quantitative values for precision, recall, and accuracy that were otherwise obscured in free-form natural language outputs.
The agentic frameworks were designed to overcome the limitations of monolithic inference by decomposing the forensic process into a series of specialized sub-tasks managed by dedicated agents. The ProvenanceAgent handled metadata and file hashes, the PerceptionAgent managed object detection and optical character recognition using YOLOv8 and Tesseract, while the MannequinAgent utilized specific visual heuristics, such as skin texture and gaze cues, to confirm human authenticity. The AttributionAgent used CLIP embeddings—tested against a gallery of five exemplars per country and branch—to ground findings in evidence, while the CrossCheckAgent was empowered to reject inconsistent reports. This modular design transformed the pipeline into a verifiable chain of custody, where each step could be audited.
When comparing performance metrics, the single-pass baseline models showed a notable imbalance, characterized by high recall but moderate precision. For instance, Gemini 1.5 Pro achieved a recall of 0.99, yet the group averaged a precision of only 0.68. The models frequently mistook mannequins for personnel, as evidenced by misclassification rates exceeding 85 percent for all baseline models. While Gemini 1.5 Pro performed best at identifying contextual attributes like country, with an accuracy of 0.787, the models consistently struggled with unit-level recognition, with even the top performer reaching only 0.254. These results confirm that unconstrained multimodal reasoning prioritizes general uniform patterns over the semantic rigor required for evidentiary standards.
The agentic integration significantly altered these metrics by introducing threshold-based logic, such as the mannequin confidence requirement of 0.6 to 0.7, and strict schema validation for all intermediate steps. The LangGraph configuration, which relied on a deterministic directed acyclic graph, reached an overall detection precision of 0.90 and limited mannequin false positives to 8 percent, a stark improvement over the baseline average of 27 percent. The collaborative CrewAI configuration also performed well, achieving 0.88 precision and 0.12 false-positive rates. While the increased processing requirements resulted in a 2.8-fold increase in latency, this cost was viewed as an acceptable trade-off for the heightened forensic confidence and auditability these systems provided.
The reduction in total recall observed when using the agentic layers, down to 0.73 for LangGraph and 0.74 for CrewAI, represents a shift from probabilistic estimation to evidence-backed reporting. Rather than failing due to hallucinated positives, the agentic systems were configured to abstain from classification in cases of ambiguity, a feature that enhances the trustworthiness of the resulting reports in a legal or security context. Qualitative analysis further emphasized this, with agentic outputs providing clear decision traces and evidentiary citations that align with established forensic documentation standards such as NIST SP 800-101 and ISO/IEC 27,037. By moving away from reliance on opaque classifiers toward structured, evidence-driven reasoning, the study demonstrates a significant methodological advancement for AI-assisted digital forensic investigations.
Subject of Research: Multimodal Large Language Models in Digital Forensic Image Analysis
Article Title: Evaluating Multimodal LLMs for Context-Aware Forensic Image Interpretation
Article References: Fedynyshyn, T., Partyka, O., Opirskyy, I., & Konakbayev, O. (2026). Evaluating Multimodal LLMs for Context-Aware Forensic Image Interpretation. International Journal of Networked and Distributed Computing, 14(2), Article 23. https://doi.org/10.1007/s44227-026-00102-6
Image Credits: AI Generated
DOI: 10.1007/s44227-026-00102-6
Keywords: Digital Forensics, Multimodal LLMs, Agentic AI, Image Interpretation, Computer Vision, CrewAI, LangGraph, Forensic Reliability, Contextual Classification, Mannequin Discrimination, Attribution Accuracy, Evidence-Grounded Reasoning, Audit-ready Reporting, Automated Investigation, Model Precision, Reproducibility, Gemini 1.5 Pro, GPT-4o, LLAVA
Cite Scienmag News
Denise Maddox. (November 19, 2026). Agentic Orchestration Improves Forensic Accuracy in Multimodal AI Image Interpretation. Scienmag. https://scienmag.com/agentic-orchestration-improves-forensic-accuracy-in-multimodal-ai-image-interpretation/
Denise Maddox. "Agentic Orchestration Improves Forensic Accuracy in Multimodal AI Image Interpretation." Scienmag, 19 November 2026, https://scienmag.com/agentic-orchestration-improves-forensic-accuracy-in-multimodal-ai-image-interpretation/. Accessed 19 November 2026.
Denise Maddox. "Agentic Orchestration Improves Forensic Accuracy in Multimodal AI Image Interpretation." Scienmag. November 19, 2026. https://scienmag.com/agentic-orchestration-improves-forensic-accuracy-in-multimodal-ai-image-interpretation/








