资源简介:
ARC-AGI-3 wa30 Agent Trajectories数据集包含一个由Claude Code驱动的大型语言模型(LLM)智能体在玩ARC-AGI-3游戏wa30时生成的游戏过程轨迹。该数据集是ARA(Agent-Native Research Artifact)世界模型泛化实验的一部分,智能体在游戏过程中实时构建结构化的世界模型,并在无法通过初始探索解决的关卡中咨询该模型以寻找解决方案。数据在游戏过程中持续同步,每当世界模型有新的结晶化内容时就会推送。数据集按目录组织,主要包括推理轨迹(包含逻辑、跟踪、暂存子目录)、每个子智能体的行动-推理轨迹(每个行动对应一行JSON记录并包含原因)、游戏状态快照和获胜动作脚本、会话级别的帧级记录、回合级别的回合日志、工具日志和会话笔记、ARC API记分卡、世界模型预测分类账(记录咨询内容及其置信度及验证结果)、游戏的分类账行和指标、消融研究和展示工件、会话索引以及来源信息(如智能体模型、仓库提交哈希等)。仓库使用L<n>-cleared标签标记关卡通关状态,这些快照可用于下游智能体评估实验。数据集适用于强化学习、智能体轨迹分析、游戏玩法研究、世界模型构建和LLM智能体行为研究等任务。
The ARC-AGI-3 wa30 Agent Trajectories dataset contains game process trajectories generated by a large language model (LLM) agent (driven by Claude Code with a file relay tool) while playing the ARC-AGI-3 game wa30. This dataset is part of the ARA world model generalization experiment. The agent constructs a structured world model (called Agent-Native Research Artifact) in real-time during gameplay and consults this model to find solutions in levels that cannot be resolved through initial exploration. Data is continuously synchronized during gameplay, pushed whenever the world model has new crystallized content. The dataset is organized by directories, including reasoning trajectories (with logic, tracking, and staging subdirectories), action-reasoning trajectories for each sub-agent (each action corresponds to a JSON record with reasons), game state snapshots and winning action scripts, session-level frame recordings, episode-level logs, tool logs and session notes, ARC API scorecards, world model prediction ledger (recording consultation content, confidence, and subsequent verification results), accounting rows and metrics for the game, ablation studies and demonstration artifacts, session index, and manifest information (such as agent model, repository commit hash, etc.). The repository uses L<n>-cleared tags to mark complete recorded states when each level is cleared, which are ready snapshots for downstream agent evaluation experiments. The dataset is suitable for tasks such as reinforcement learning, agent trajectory analysis, gameplay research, world model construction, and LLM agent behavior research.