AI agents can sound strategic while reasoning from events that never happened

Описание и применение

В минимально обоснованной среде ИИ-агенты создавали правдоподобные рассуждения на основе несуществующих событий, что приводило к галлюцинациям. Добавление канонического протокола, голосований и ночных действий создало наблюдаемые последствия, позволив проверять и корректировать ошибки.

Задача

AI agents produced plausible social reasoning based on nonexistent evidence, leading to hallucinations in a multi‑agent game.

Результат

Grounding the environment with a canonical record and observable consequences reduced unsupported reasoning and made factual errors contestable, though agents remained imperfect.

Рабочий процесс

• Design minimal game with no observable state
• Run three games with Hyperagent and observe hallucinations
• Add canonical public record, eliminations, night kills, and private info
• Run richer games (V2, V3) and monitor claim verifiability
• Document errors and corrections via operator notes

Доказательства

In the minimally grounded version, all three games contained confident social reasoning based on nonexistent behavior or operator-controlled conditions.

Adding a public transcript, eliminations, night actions, and private information created observable consequences and a shared record, allowing errors to be contested.