TTT-Bench
收藏资源简介:
TTT-Bench是一个用于评估大型推理模型(LRMs)在基本战略、空间和逻辑推理能力方面的新基准,它通过一系列四个人工设计的两人制井字游戏来测试这些能力。这些游戏虽然对人类来说很简单,但需要推理对手的意图以及游戏板的空间配置。数据集包含了四种游戏类型:普通井字游戏(oTTT)、双重井字游戏(dTTT)、立方体井字游戏(cTTT)和正方形井字游戏(sTTT)。每个游戏都有其独特的胜利条件和游戏状态。数据集包含约100个样本,每个游戏类型都有可验证的答案。TTT-Bench旨在评估LRMs在解决需要基本战略、空间和逻辑推理的简单任务方面的能力,以解决当前LRMs在数学以外的推理能力评估不足的问题。
TTT-Bench is a novel benchmark for evaluating Large Reasoning Models (LRMs) on their fundamental strategic, spatial, and logical reasoning capabilities. It tests these abilities via a set of four manually designed two-player tic-tac-toe games. Though simple for humans, these games require reasoning about opponents' intentions and the spatial configuration of the game board. The dataset includes four game variants: ordinary tic-tac-toe (oTTT), double tic-tac-toe (dTTT), cubic tic-tac-toe (cTTT), and square tic-tac-toe (sTTT). Each variant features unique winning conditions and game states. The dataset contains roughly 100 samples, with verifiable answers available for each game type. TTT-Bench is designed to assess LRMs' ability to solve simple tasks demanding basic strategic, spatial, and logical reasoning, thereby addressing the current gap in evaluating LRMs' reasoning capabilities beyond mathematical contexts.

- 1TTT-Bench: A Benchmark for Evaluating Reasoning Ability with Simple and Novel Tic-Tac-Toe-style GamesAMD(美国超威半导体公司) · 2025年



