Safety of Latent Communication in Multi-Agent Systems

DOI10.48550/arXiv.2609.39788MaRDI QIDQ7361999arXivFDO

Authors Muhammad Huzaifa Yaseen, Thorsten Eisenhofer, Muhammad Huzaifa, Sina Mavali

Publication date 30 September 2026

Published in arXiv

Tldr We study latent communication in multi-agent systems and find that benignly trained links can increase harmful compliance compared to text-based communication. A reinforcement-learning attack that rewards both harmless response accuracy and harmful compliance further amplifies this effect, raising the mean harmful-compliance score from 27.9 to 76.9 across three topologies and four safety benchmarks.

Full work available at URL https://doi.org/10.48550/arXiv.2609.39788

arXiv classification

cs.AIcs.LGcs.MA

Cites work

• Decoupled Weight Decay Regularization

This page was built for publication: Safety of Latent Communication in Multi-Agent Systems

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7361999)