EVENT DOSSIER
The researchers proposed a zero-sample pipeline called Abstraction Agent, which uses large language models to extract continuous strategic features from natural language game descriptions and clusters private states into abstract buckets. This process consists of four stages: feature discovery, state scoring, feature selection, and k-means clustering, without the need for a specific evaluator or training data. In head-up no-limit Texas hold'em (HUNL) final-hand tests, this method increased strategy utilization by up to 62% and surpassed the scalar ranking baseline in ROVER Trials. Additionally, this pipeline can be migrated to four games—four-card limit Omaha, HUNL pre-flop and flop stages, and Riichi Mahjong—without modifying the prompts. The features discovered by this method tracked the recognized strategic concepts of each game. The relevant code is available on GitHub.
提出名为 Abstraction Agent 的零样本流水线，利用大语言模型从自然语言游戏描述中挖掘连续战略特征、对私有状态进行评分并聚类为抽象桶，无需特定评估器或训练数据。该流程包含四个阶段：带校准锚点的特征发现、批处理私有状态评分、基于相关性的特征选择及 k-means 聚类。在heads-up no-limit Texas hold'em (HUNL) 转牌终局测试中，生成的抽象将提升策略的利用性降低高达 62%，并在 ROVER Trials 上超越标量排名基准。此外，该管道未修改提示词即可迁移至四张牌限注奥马哈、HUNL 翻前及翻牌阶段以及 Riichi Mahjong 四种游戏，其发现的特征追踪了各游戏的公认战略概念。代码已公开于 GitHub。