Decrypto Benchmark
收藏资源简介:
Decrypto是一个基于语言游戏的多人推理和心智理论(ToM)基准,由FAIR at Meta和牛津大学的研究人员开发。该数据集基于Scorpion Masqué公司发布的同名桌面游戏,旨在通过多人互动游戏来评估大型语言模型在推理和ToM方面的能力。数据集包含680个可能的关键词,超过8.8亿种可能的组合,以及无限数量的游戏可能性,因为提示的选择是开放式的,可以依赖于任何公开可用的信息。该数据集可用于评估LLM在竞争和合作场景下的推理能力,以及在进行战略决策前对他人可获得的信息进行推理的能力。
Decrypto is a language-game-based multi-player reasoning and Theory of Mind (ToM) benchmark developed by researchers at FAIR at Meta and the University of Oxford. This dataset is based on the eponymous tabletop game published by Scorpion Masqué, and aims to evaluate the reasoning and ToM capabilities of large language models (LLMs) via multi-player interactive games. The dataset contains 680 possible keywords, over 880 million possible combinations, and an unlimited number of game possibilities, as the selection of prompts is open-ended and can rely on any publicly available information. This dataset can be used to evaluate LLMs' reasoning abilities in both competitive and collaborative scenarios, as well as their ability to reason about the information accessible to others before making strategic decisions.

- 1The Decrypto Benchmark for Multi-Agent Reasoning and Theory of MindFAIR at Meta, University of Oxford · 2025年



