Outbidding and Outbluffing Elite Humans: Mastering Liar's Poker via Self-Play and Reinforcement Learning

Researchers have developed Solly, an AI agent that achieved elite human-level performance in Liar's Poker through self-play reinforcement learning, winning over 50% of hands against top players. This breakthrough extends AI capabilities beyond two-player games to complex multi-player scenarios with imperfect information, demonstrating novel strategic behaviors that resist exploitation by world-class competitors.

Solly's achievement represents a meaningful advance in multi-agent AI systems operating under uncertainty and imperfect information. Unlike previous poker AI breakthroughs that primarily focused on two-player no-limit Texas hold'em—where hands typically resolve quickly—Liar's Poker demands sustained multi-player engagement across extended bidding sequences, requiring more sophisticated decision-making frameworks. The use of model-free actor-critic deep reinforcement learning proves effective for discovering non-obvious strategies that human experts struggle to counter.

This work builds on decades of game-theoretic AI research, where imperfect-information games serve as rigorous testing grounds for reasoning under uncertainty. Texas hold'em victories demonstrated AI could match human intuition in specific domains; Solly extends this to environments where strategic depth scales with player count. The finding that large language models underperformed reinforcement-learning agents on identical metrics suggests that scale and linguistic capability alone cannot solve strategic reasoning tasks requiring probabilistic reasoning and adversarial adaptation.

The implications extend beyond poker. Multi-player games with imperfect information model real-world scenarios—negotiation, auction bidding, resource allocation—where groups make decisions with asymmetric information. AI systems capable of elite performance in these settings could inform optimization in business and economic contexts. However, the current contribution remains academically focused; practical applications depend on translating poker-specific strategies to domain-specific problems.

• →Solly achieved elite human-level play in multi-player Liar's Poker using deep reinforcement learning self-play, winning over 50% of hands.
• →The AI developed novel, randomized bidding strategies that world-class human players found difficult to exploit or predict.
• →Large language models, including reasoning-capable variants, significantly underperformed the reinforcement learning agent on identical competitive metrics.
• →The breakthrough extends AI capabilities beyond two-player games to complex multi-player scenarios requiring sustained engagement and imperfect information handling.
• →Multi-player game mastery demonstrates AI progress toward strategic reasoning applicable to real-world negotiation and bidding scenarios.