IJCCI_DC 2026 Abstracts

Short Papers
Short Papers
Paper Nr: 6
Title:

Efficient Architecture Evaluation and Decision Making in Multi-Objective Neural Architecture Search for Super-Resolution
Authors:

Sergio Sarmiento-Rosales
Abstract: Neural architecture search (NAS) can discover super resolution networks that balance reconstruction quality and computa tional complexity, but its usefulness is limited by the cost and uncertainty of candidate evaluation. This doctoral research studies that bottleneck as a sequence of resource-allocation decisions: how to estimate quality during search, which estimator is appropriate for a given stage, and which predicted candidates justify full training. The work progressed from an XGBoost surrogate, through a controlled comparison of learned predictors and SynFlow, to a systematic benchmark of model-based predictors. Those studies showed that online scoring cost cannot be interpreted without the offline cost of acquiring training data, and that no evaluator simultaneously dominates ranking fidelity, search diversity, final trained outcomes, and cost. Current work consequently examines two complementary layers. First, offline Pareto archiving reduces a large predicted set to a diverse validation shortlist; this is a post-search decision mechanism, not a predictor. Second, eleven zero-cost proxies are benchmarked under a common super-resolution protocol, with explicit parameter-count baselines and size-confounding diagnostics. In the current Strategic-50 sample, parameters and fixed-resolution convolutional FLOPs are perfectly rank-aligned, supporting one active complexity objective rather than two in recent Bass searches. The ex pected thesis contribution is an evidence-based framework for selecting and validating architecture-evaluation strategies across broad screening, multi-objective search, Pareto-conditioned reranking, and final training. Remaining work concerns active integration of promising zero-cost signals and external validation beyond one search space, dataset, and scale factor.
Paper Nr: 10
Title:

Evidence-Centered XAI for Trustworthy and Adaptive Industrial Visual Inspection
Authors:

Abdullah Abdullah
Abstract: Deep-learning-based visual inspection is increasingly used in industrial quality control, yet high benchmark accuracy does not ensure that an individual decision is trustworthy. A detector may be confidently wrong, rely on misleading contextual cues, produce plausible but unfaithful explanations, or degrade when materials, sensors, illumination, or defect distributions change. Explainable AI (XAI) is therefore central to this doctoral research, not merely as a visualization layer but as a source of measurable evidence for diagnosing what drives predictions, where model reliance is misplaced, and when confidence may be misleading. The research develops an evidence-centered framework that connects XAI with predictive calibration, uncertainty, robustness, and human oversight at the level of individual detections. A preliminary steel-defect study motivates this direction by showing that localization and faithfulness provide different assessments of explanation quality across defect morphologies. Building on this foundation, the doctoral programme will examine how predictive and explanatory evidence interact, how they change under perturbation and distribution shift, and whether their combination can reveal unreliable detections beyond confidence alone. Diagnosed failures will further support actionable XAI for model adaptation, selective human-in-the-loop action, and evidence-grounded visual and natural-language explanations that communicate model evidence and limitations. The expected outcome is a lifecycle framework for explainable, trustworthy, adaptive, and human-centered industrial visual inspection.
Paper Nr: 11
Title:

MEDICAE: Personalized, Explainable and Proactive Multi-Agent Assistance for Telemedicine Consultations
Authors:

Samir Jabbar
Abstract: Agents built on large language models (LLMs) can now hold clinical conversations, but in a regulated medical setting the capabilities that would make them useful beyond answering questions each pull against medical validity. Adapting a recommendation to a patient can change its clinical content. An explanation can be plausible without reflecting what actually drove a decision. An agent that intervenes unprompted can intrude on or alarm the patient. This doctoral project makes that constraint, rather than raw capability, its object of study. It asks how a multi-agent system can personalize recommendations to a patient's clinical, cognitive and affective state without altering their validity (RQ1), how it can explain its decisions faithfully and attribute them to the agents that actually determined them (RQ2), and how, building on both, it can decide on its own initiative whether, when and how to intervene (RQ3). Each question is paired with a validity criterion that the architecture enforces and the evaluation measures: clinical invariance for personalization, trace-grounded faithfulness and attribution for explanation, and safety and appropriateness constraints for initiative. The approach is embodied in MEDICAE, a domain-general multi-agent architecture in which plans and agent roles are generated at runtime under supervisor checkpoints, clinical judgement is made by models reasoning over a typed and auditable patient state, and the safety-critical parts of delivery stay deterministic. A first system, DocAgent-XAI, shows that our method achieves the lowest diagnostic error rate (4.3%) and the highest clinical quality scores (LLM-as-Judge: 2.3/3). MEDICAE is implemented, with a first use case in cardiology. Its evaluation with practitioners, patients and public conversational datasets is the next step.
Paper Nr: 11
Title:

MEDICAE: Personalized, Explainable and Proactive Multi-Agent Assistance for Telemedicine Consultations
Authors:

Samir Jabbar
Abstract: Agents built on large language models (LLMs) can now hold clinical conversations, but in a regulated medical setting the capabilities that would make them useful beyond answering questions each pull against medical validity. Adapting a recommendation to a patient can change its clinical content. An explanation can be plausible without reflecting what actually drove a decision. An agent that intervenes unprompted can intrude on or alarm the patient. This doctoral project makes that constraint, rather than raw capability, its object of study. It asks how a multi-agent system can personalize recommendations to a patient's clinical, cognitive and affective state without altering their validity (RQ1), how it can explain its decisions faithfully and attribute them to the agents that actually determined them (RQ2), and how, building on both, it can decide on its own initiative whether, when and how to intervene (RQ3). Each question is paired with a validity criterion that the architecture enforces and the evaluation measures: clinical invariance for personalization, trace-grounded faithfulness and attribution for explanation, and safety and appropriateness constraints for initiative. The approach is embodied in MEDICAE, a domain-general multi-agent architecture in which plans and agent roles are generated at runtime under supervisor checkpoints, clinical judgement is made by models reasoning over a typed and auditable patient state, and the safety-critical parts of delivery stay deterministic. A first system, DocAgent-XAI, shows that our method achieves the lowest diagnostic error rate (4.3%) and the highest clinical quality scores (LLM-as-Judge: 2.3/3). MEDICAE is implemented, with a first use case in cardiology. Its evaluation with practitioners, patients and public conversational datasets is the next step.