QUART: Agentic Reasoning To Discover Missing Knowledge in Multi-Domain Temporal Data.

ICLR 2026 Conference Submission9627 Authors

Keywords: Agentic AI, Missing Information, Temporal Reasoning, Multi-Domain Learning, Reinforcement Learning, Neural-Symbolic, Expert Systems, Causal Inference

Abstract:

Existing AI systems for expert decision support, commonly treat incomplete information as missing data to be filled or ignored, but this essentially misapprehends the fundamental challenge experts face: recognizing what crucial information is still unknown. Methods like knowledge graphs and LSTMs rely on temporal patterns and embeddings, which reduce interpretability and fail to address knowledge
gaps through logical relations. We present QUART (Query-based Understanding Agent for Reasoning Temporal data), a multi-domain framework that preserves semantic meaning and actively learns to identify and recall unexamined information by reasoning over causal connections via reinforcement learning. QUART integrates interpretable semantic causal graphs, multi-agent policies optimizing
decision utility, and explicit modeling of unknown information to drive strategic questioning based on logical causal dependencies instead of temporal sequences. It encodes clinical outputs dynamically, approximating real expert workflows with confidence measurements. Evaluated on over 40,000 patient histories from the MIMIC-III healthcare database, QUART achieves over 10% higher diagnostic accuracy than LSTM baselines, not by imputing missing data but by revealing and addressing information blind spots. Furthermore, its dynamic agents protect privacy by querying sensitive data only when necessary. Although illustrated on medical data, pilot studies are underway exploring the framework’s potential in education, product management, and office decision-making. This work lays a
foundation for trustworthy, interpretable AI assistants that improve expert decision-making in complex and sensitive environments.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 9627