Large language models (LLMs), the computational models that underpin conversational agents such as Gemini and ChatGPT, are now broadly used by people worldwide to quickly find information, summarize documents and generate texts for particular purposes. Some computer scientists at the moment are integrating two or more of these models to form multi-agent systems, which prompt numerous artificial intelligence (AI) agents to engage, cooperate and/or compete with the purpose of completing particular tasks.
In a few scenarios, however, AI gents ought to have different objectives and might have access to more or less information than the other agents they’re interacting with. Understanding how AI agents usually behave in these situations could support shed more light on the potential advantages and risks of multi-agent systems.
Researchers at Mila, Université de Montréal and McGill University lately set out to discover how the hidden aims of character AI agents could influence a multi-agent system’s performance, using a framework inspired by the multiplayer social deduction game Werewolf.
The outcomes of this AI social experiment, published on the arXiv preprint server, disclosed that when even a single AI agent had slightly different objectives from those of its peers, it could poorly have an impact on the overall performance of an whole multi-agent system.
Testing LLM-based agents in a social deduction game
To explore how conflicting AI agent objectives affected the performance of multi-agent systems, the researchers tested LLM-based agents in a game referred to as Werewolf. This is a famous social deduction game that prompts players to talk about and examine one another to decide who’s secretly working in against the group.
“LLM-powered multi-agent systems are increasingly more deployed in integrated-motive environments, wherein agents operate under asymmetric information and strategic deception due to conflicting or hidden goals,” Marylou Fauchard, Florian Carichon and their colleagues write of their paper.
“In these settings, misalignment with collective goals turns into a central concern. We propose a novel framework for analyzing objective misalignment the use of the social deduction game Werewolf, modifying the goal of a single agent when preserving its assigned function.”
In the Werewolf game, each player is assigned a particular role that isn’t known to other players. Most players are villagers, and their aim is to detect and eliminate werewolves. A smaller number of gamers are werewolves. These players know who the other werewolves are, and their purpose is to secretly do eliminate villagers without being discovered.
A few other players can take on the role of seer or doctor. Seers can secretly learn the role of another player every turn, while doctors can save one player from elimination.
Every night, werewolves eliminate one player. During the day, all remaining players vote to eliminate a suspected werewolf, while werewolves normally try to persuade the group to eliminate someone they know isn’t always a werewolf instead. The game ends either when all werewolves have been eliminated or after they reach parity (i.E., when there are as many werewolves as villagers).
Fauchard, Carichon and their colleagues tested a group of AI agents in this game, assigning one of the 4 roles to each of them and then observing how they interacted with each other. They also looked at how secretly changing the goal of a single AI agent affected its reasoning, communication and the overall outcome of the game.
“Across LLMs from four different model families and sizes, 4 player roles, and three objective formulations, we introduce a dual analysis of the agents’ internal reasoning and their public cheap-talk behavior (i.e., costless, non-binding communication that does not directly affect the agent’s utilities), complemented by means of an evaluation of game outcomes,” write the authors.
“Our outcomes show that objective misalignment undermines outcomes in inherently adversarial environments, an effect exacerbated via asymmetric information and specialised roles. While compromised agents continuously broaden distinct objective-dependent reasoning strategies, those adaptations stay to be largely invisible in their public behavior.”
Informing the future design of multi-agent systems
The outcomes of this study emphasize a potential limitation of multi-agent systems. Particularly, they demonstrate that if interacting agents have different objectives, their actions could influence the overall performance of a whole multi-agent system on a given task. In addition, the researchers discovered that the altered objectives of individual agents were often evident from their internal reasoning, but they were tough to infer from their communications with other agents.
“Our findings suggest that even subtle objective misalignment can profoundly have an effect collective decision-making, emphasizing the need for effective mitigation strategies for LLM-based multi-agent system,” the authors write.
In the future, the insights gathered by this research team could potentially support the development of safer and more dependable multi-agent AI system that prompt individual agents to prioritize a shared goal.
