CANDICE: Agentic Causal Disentanglement with Class Conditional Knowledge Integration for Long Tailed Domain Generalization

ACL ARR 2026 January Submission10645 Authors

Keywords: Agentic language models, Causal reasoning, Class-conditional knowledge integration, Domain generalization, Long-tailed learning, Neurosymbolic NLP, Tool-grounded reasoning, Robust inference

Abstract: Deep learning models deployed in clinical set- tings face two main challenges: Domain Gen- eralization (DG) and Long Tailed Visual Recognition (LT VR). DG mandates learn- ing domain invariant features to perform ro- bustly across heterogeneous acquisition pro- tocols and patient populations. We formally investigate the theoretical trade off where gradient alignment objectives prioritized by DG methods undermine the class aware optimization necessary for LT VR . To resolve this issue, we introduce the Agentic Causal Disentangle- ment (CANDICE) Framework, a novel modular architecture that integrates explicit clinical knowledge/expertise sourced from sonographers, radiologists, and specialists as a causal intervention tool autonomously. CANDICE orchestrates three specialized agents: the Clinical Reasoning Agent (CRA), the Causal Dis- entanglement Agent (CDA), and the Code Generation Agent (CGA). This integration of domain specific, causal knowledge effec- tively decouples the DG and LT objectives. Evaluated across 10 diverse medical imaging datasets spanning 4 modalities, the CANDICE Framework overall achieves an average 10.3% performance improvement across multi domain and in domain long tailed tasks.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: Agentic language models, Causal reasoning, Class-conditional knowledge integration, Domain generalization, Long-tailed learning, Neurosymbolic NLP, Tool-grounded reasoning, Robust inference

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 10645