Even More Deception: Objective Misalignment in Mixed-Motive LLM Multi-Agent Systems

DOI10.48550/arXiv.2607.26120MaRDI QIDQ7284615arXivFDO

Authors Margarida Carvalho, Golnoosh Farnadi, Marylou Fauchard, Florian Carichon

Publication date 28 July 2026

Published in arXiv

Tldr The paper introduces a novel framework to evaluate objective misalignment in large language model (LLM) multi-agent systems using the social deduction game Werewolf and finds that even subtle objective misalignment can significantly impact collective decision-making outcomes, with compromised agents developing distinct reasoning strategies but keeping them hidden from public behavior.

Full work available at URL https://doi.org/10.48550/arXiv.2607.26120

arXiv classification

cs.AI

Cites work

• Mafia: A theoretical study of players and coalitions in a partial information environment
• UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction
• When to adjust alpha during multiple testing: A consideration of disjunction, conjunction, and individual testing

This page was built for publication: Even More Deception: Objective Misalignment in Mixed-Motive LLM Multi-Agent Systems

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7284615)