What is it about?
Multi-agent LLMs can examine an utterance from different perspectives, but repeating this process for every input takes considerable time. REMICA builds a memory of multi-agent predictions and rationales in advance. For a new utterance, it retrieves similar records from this memory and produces a prediction with a single LLM call. The memory is refined using the correct labels during offline construction, while no labels are used at test time. Experiments on four datasets suggest that reusing multi-agent reasoning in this way can substantially reduce inference time and help improve the efficiency of inappropriate utterance detection.
Featured Image
Photo by Shrinath on Unsplash
Why is it important?
The cost and latency of repeated LLM interactions can limit the practical use of multi-agent systems. REMICA moves multi-agent interactions to an offline stage and reuses the resulting reasoning records during inference. In the best matched comparison, it was 7.80 times faster than online multi-agent inference and also achieved a higher average F1 score. This suggests that multi-agent reasoning can be reused more efficiently in content moderation and online safety applications.
Perspectives
I wanted to explore whether, for a new utterance, the past judgments and rationales of multiple agents on similar, rather than identical, examples could provide more useful guidance than the text-label pairs used in a standard RAG baseline. This question led us to treat previous multi-agent reasoning as reusable memory. The study also made me consider the quality of that memory, since inaccurate rationales may otherwise be retrieved and influence later predictions.
Juyoung Kim
Gyeongsang National University
Read the Original
This page is a summary of: REMICA: Reflective Memory and Interventional Context Alignment with Multi-Agent LLMs for Inappropriate Utterance Detection, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3805712.3809888.
You can read the full text:
Contributors
The following have contributed to this page
