ALMMIT: Agent-Guided Active Learning for Multimodal Misinformation Detection on Temporal Graph Data

ACL ARR 2026 January Submission10176 Authors

Keywords: misinformation detection, agent-guided active learning, multimodal, reinforcement learning, temporal graph, modality-aware loss

Abstract: Posting information on a social platform is effortless, often just a single click away; yet, fact-checking mechanisms remain limited, and authenticating information is challenging due to the dynamism of misinformation propagation. The existing misinformation detection frameworks over-rely on annotated data and fall short in managing various modalities (Text, Images, conversational structure) that contribute to misinformation evolution. To mitigate this challenge, we propose ALMMIT, an Active Learning multimodal framework for Misinformation Detection, guided by a Reinforcement Learning (RL) agent. The RL agent adaptively directs the active learning process, maintaining momentum across rounds and avoiding bias towards the majority class while operating effectively in few-label and semi-supervised settings to overcome data scarcity within an imbalanced dataset. In addition, ALMMIT employs a modality-aware loss to dynamically adjust modality weights during fine-tuning and active learning, and introduces a composite reward mechanism that balances momentum gain with targeted attention to minority classes. Experimental results show that ALMMIT consistently outperforms recently developed state-of-the-art few-label and supervised approaches.

Paper Type: Long

Research Area: Low-resource Methods for NLP

Research Area Keywords: misinformation detection, agent-guided active learning, multimodal, reinforcement learning, temporal graph, modality-aware loss

Contribution Types: Approaches to low-resource settings

Languages Studied: English

Submission Number: 10176