parsa-mz/puzzlezoo
收藏资源简介:
PuzzleZoo是一个包含三个推理/规划基准的数据集,旨在评估大型语言模型在多步骤程序性问题解决上的性能。该数据集是论文《RePoT: Recoverable Program-of-Thought via Checkpoint Repair》的官方评估套件,并作为独立基准发布给更广泛的社区。数据集包括:1) PuzzleZoo-775:包含775个分层谜题实例,覆盖四个经典规划环境(汉诺塔、跳棋过河、河内塔、积木世界),每个环境在可控复杂性下生成,以测量解决率随规划长度下降的情况;2) PlanBench-Blocksworld-378:包含378个Blocksworld实例(3-12个积木),改编自PlanBench,采用与PuzzleZoo相同的记录模式,用于第三方复制;3) Derail-550:包含550个方法无关的中途错误注入案例,每个案例固定一个问题,部分执行预言计划,注入单个错误动作,并将结果状态传递给恢复方法,以隔离驱动恢复的信号。总计1703条记录,所有配置模式兼容。该基准的特点包括按复杂性分层、提供步骤级验证器(无LLM判断)、以及方法无关的恢复评估,支持多步骤程序规划、错误恢复能力和方法对检查点信号依赖的评估。
PuzzleZoo is a collection of three reasoning/planning benchmarks designed to evaluate large language models on multi-step procedural problem-solving. It is the official evaluation suite for the paper RePoT: Recoverable Program-of-Thought via Checkpoint Repair and is released as a standalone benchmark for the broader community. The dataset includes: 1) PuzzleZoo-775 — 775 stratified puzzle instances across four classical planning environments (Tower of Hanoi, Checker Jumping, River Crossing, Blocksworld), each generated at controllable complexity to measure how solve-rate degrades with horizon length; 2) PlanBench-Blocksworld-378 — 378 Blocksworld instances (3–12 blocks) adapted from PlanBench into the same record schema as PuzzleZoo, used as a third-party replication target; 3) Derail-550 — 550 method-agnostic mid-rollout error-injection cases, each fixing a problem, walking an oracle plan partway, injecting a single wrong action, and handing the resulting state to a recovery method to isolate which signal drives recovery. Together, the three configs cover 1703 records, all schema-compatible. The benchmark features stratification by complexity, step-level verifier (no LLM judge), and method-agnostic recovery evaluation, intended to assess multi-step procedural planning, error recovery/replanning ability, and how a methods effectiveness depends on checkpoint signals.




