AI Agent Solly Masters Liar's Poker at Elite Human Level
A team of researchers has introduced Solly, the inaugural AI agent capable of reaching elite human-level performance in reduced-format Liar's Poker, a game known for its intricate multi-player interactions and incomplete information. Utilizing a model-free, actor-critic deep reinforcement learning algorithm, Solly was trained through self-play. It secured victories in over 50% of hands and exhibited exceptional equity in both heads-up and multi-player scenarios. Furthermore, Solly surpassed large language models (LLMs), including those proficient in reasoning, based on the same evaluation criteria. This advancement marks a significant expansion of AI capabilities beyond Texas hold'em, venturing into games with more intricate multi-player dynamics.
Key facts
- Solly is the first AI agent to achieve elite human play in reduced-format Liar's Poker.
- Solly was trained using self-play with a model-free, actor-critic, deep reinforcement learning algorithm.
- Solly won over 50% of hands in heads-up and multi-player Liar's Poker.
- Solly outperformed large language models (LLMs) including those with reasoning abilities.
- The research was published on arXiv with ID 2511.03724v3.
- Liar's Poker involves extensive multi-player engagement and imperfect information.
- Previous AI breakthroughs focused on no-limit Texas hold'em with subdued multi-player dynamics.
- Solly uses a model-free approach without explicit game rules.
Entities
Institutions
- arXiv