REMICA: Reflective Memory and Interventional Context Alignment with Multi-Agent LLMs for Inappropriate Utterance Detection

초록

Recently, multi-agent debate-based inference methods have been widely adopted for inappropriate utterance detection to aggregate diverse perspectives and improve explainability. However, performing real-time interactions for every input incurs substantial latency and cost, and may amplify error entrenchment as early misjudgments are repeatedly justified. In this paper, we propose the Reflective Memory and Interventional Context Alignment (REMICA) framework. Specifically, REM stores multi-agent predictions and rationales as offline memory for efficient retrieval-based inference. During this process, ICA enforces alignment constraints to ensure consistency between rationales and predictions. Experiments on four datasets with three commercial LLMs show that REMICA reduces the average inference time relative to online multi-agent inference while improving the average F1 score, with GPT-5.1-mini achieving a relative average F1 improvement of +2.594% and an average inference-time reduction of 7.80× in the best-performing setting. These results suggest that the proposed framework can help mitigate the trade-off between accuracy and latency by reusing offline memory and reducing the risk of error entrenchment. Furthermore, the memory and code developed in this paper are publicly available at https://github.com/wndudwkd003/REMICA. © 2026 Owner/Author.

키워드