AI

二者間情報非対称性を持つ文脈バンディット監視ゲーム

A Contextual-Bandit Oversight Game with Two-Sided Informational Asymmetry

arxiv2026年7月2日

日本語要約

本稿は、二者間情報非対称性を持つ文脈バンディット監視ゲームを提案する。監視者とエージェントが異なる情報を持つシナリオをモデル化し、効果的なAI監視戦略の開発を目指す。これは、複雑で情報が乏しい環境でのAIの安全性と制御を確保する上で重要である。

English Summary

This paper proposes a contextual-bandit oversight game with two-sided informational asymmetry. It models a scenario where an overseer and an agent have different information, aiming to develop effective AI oversight strategies. This has significant implications for ensuring AI safety and control in complex, information-scarce environments.


元記事を読む𝕏 でシェア