arc-agi3-vc33-agent-trajectories
收藏资源简介:
ARC-AGI-3 vc33 Agent Trajectories 是一个开源数据集,专注于记录基于Claude Code驱动的LLM智能体在玩ARC-AGI-3游戏vc33版本时的游戏轨迹。该数据集是ARA-as-world-model泛化实验的一部分,旨在探索智能体如何通过构建和利用结构化世界模型(称为Agent-Native Research Artifact)来提升游戏表现。智能体在实时游戏过程中动态建立世界模型,并在遇到无法仅通过初始探索解决的关卡时参考该模型以破解难题。数据在游戏过程中持续同步,每次世界模型的固化都会更新到数据集中。数据集内容涵盖多个方面:包括智能体的世界模型和推理痕迹(位于reasoning/ara-vc33/目录)、每个子智能体的动作-推理轨迹(每行动作对应一条JSON行并附带理由说明)、解决方案状态快照(记录进度、机制、陷阱和暂停原因)、每关获胜动作脚本、帧级录制文件(每个会话一个文件)、回合级事件日志(跨运行拼接)、工具日志和自由格式会话笔记、ARC API计分卡、世界模型预测账本(记录咨询信心和后续验证)、游戏账目记录(如代币/步骤指标、ARA增长曲线)、演示工件(用于消融或展示),以及会话索引和清单文件(提供数据来源信息,如智能体模型、代码库提交、协议哈希和同步时间)。数据集通过repo标签L<n>-cleared标记每个关卡清除时的完整记录快照,这些快照适用于下游智能体评估实验,作为保持独立的基准。此外,审议时间间隔、死亡分析和惊喜分析等衍生信息可以从录制文件和轨迹中计算得出,无需单独存储。该数据集适用于强化学习、游戏玩法分析、LLM智能体行为研究、世界模型构建与泛化实验等任务,为研究人员提供了详细的智能体决策过程和交互轨迹数据。
ARC-AGI-3 vc33 Agent Trajectories is an open-source dataset dedicated to recording the gameplay trajectories of LLM-powered agents (which utilize Claude Code as their file-relay tool) while playing the ARC-AGI-3 game vc33. This dataset is part of the ARA-as-world-model generalization experiments, aiming to explore how agents enhance their gameplay performance by constructing and leveraging structured world models termed Agent-Native Research Artifacts. Agents dynamically build world models during real-time gameplay, and refer to these models to solve puzzles that cannot be resolved solely through initial exploration. The dataset is synchronized in real time during gameplay, with each finalized world model updating the dataset. The dataset encompasses multiple components: the agent's world models and reasoning traces (located in the reasoning/ara-vc33/ directory), action-reasoning trajectories of each sub-agent (each action corresponds to a JSON line with accompanying justifications), solution state snapshots (recording progress, mechanisms, traps, and pause reasons), winning action scripts for each level, frame-level recording files (one file per session), round-level event logs (stitched across multiple runs), tool logs and free-form session notes, ARC API scorecards, world model prediction ledgers (recording consultation confidence and subsequent validation), game accounting records (such as token/step metrics and ARA growth curves), demonstration artifacts (for ablation studies or demonstrations), as well as session indices and manifest files that provide data source information including agent model, codebase commit, protocol hash, and synchronization time. The dataset marks complete record snapshots upon each level clearance via the repo tag L<n>-cleared; these snapshots serve as independent baselines for downstream agent evaluation experiments. Additionally, derivative information such as deliberation time intervals, death analysis, and surprise analysis can be derived from the recording files and trajectories without requiring separate storage. This dataset is applicable to tasks including reinforcement learning, gameplay analysis, LLM agent behavior research, world model construction and generalization experiments, providing researchers with detailed data on agent decision-making processes and interactive trajectories.
数据集概述
- 数据集名称: ARC-AGI-3
vc33— Agent Trajectories - 许可证: CC-BY-4.0
- 任务类别: 强化学习、其他
- 语言: 英语
- 标签: arc-agi、arc-agi-3、agent-trajectories、game-playing、llm-agent、world-model、claude
- 数据集别名: ARC-AGI-3 vc33 Agent Trajectories
数据集描述
该数据集包含一个LLM智能体(使用Claude Code驱动文件中继框架)在玩ARC-AGI-3游戏vc33时的游戏轨迹。该游戏属于“ARA作为世界模型”泛化实验的一部分。智能体在游戏过程中实时构建结构化世界模型(一种智能体原生研究工件),并在运行过程中参考该模型来攻克通过冷探索无法解决的关卡。该数据集还有一个配套数据集:arc-agi3-ls20-agent-trajectories。
在游戏过程中,数据会持续同步——每次世界模型结晶都会触发数据推送。
数据集布局
reasoning/ara-vc33/ # 智能体的世界模型及推理轨迹(logic/、trace/、staging/) traces/ # 每个子智能体的行为-推理轨迹:每行为一个JSON对象,包含行为及其原因 solutions/GAME.md # 状态快照:进度/机制/陷阱/暂停原因 solutions/replay/ # 每关获胜行为脚本(L<n>.txt) recordings/ # 逐帧记录,每会话一个文件(<game>-<hash>.<agent>.<guid>.recording.jsonl) episodes/ # 回合级剧集日志(跨运行拼接而成) logs/ # 中继框架日志 + 自由格式会话笔记(git历史保留之前的会话) scorecards/ # ARC API得分卡(如有保存) predictions.jsonl # 世界模型预测账本:查询记录(含置信度及后续判定结果) accounting/ # 该游戏的账本行(token/步骤数)、指标部分、ARA增长曲线 demo/ # 消融/展示工件(如有) sessions_index.json # 每个录制会话的条目(智能体、guid、帧数、最终状态) manifest.json # 溯源信息:智能体模型、仓库提交、协议哈希、同步时间
额外说明
- 仓库标签
L<n>-cleared标记每个关卡通关时的完整记录,可作为下游智能体评估实验的保持就绪快照。 - 推理时机(行为间隔)、死亡原因分析及意外分析可从记录和轨迹中推导得出,未单独存储。




