arc-agi3-agy-gemini3.1pro-ls20
收藏资源简介：
ARC-AGI-3 ls20 Agent Trajectories数据集记录了agy-gemini3.1pro模型在ARC-AGI-3游戏ls20中的完整游戏轨迹，是ARA作为世界模型泛化实验的组成部分。该数据集的核心特征是智能体在游戏过程中实时构建结构化世界模型（称为Agent-Native Research Artifact），并利用该模型解决那些仅通过初始探索无法破解的关卡。数据集采用每个harness×模型×游戏组合独立存储的结构，相同游戏由不同模型对局的数据分布在兄弟仓库中，便于比较分析。数据在游戏过程中持续同步，每次世界模型的固化都会即时推送更新。数据集包含多个维度的记录：推理痕迹目录保存智能体的世界模型构建过程和逻辑推理轨迹；行动轨迹目录以JSON行格式记录每个子智能体的行动及决策依据；解决方案目录包含关卡进度快照和获胜行动脚本；录制文件提供帧级游戏记录；此外还包括回合级日志、工具日志、API评分卡、世界模型预测账本、资源消耗统计等丰富内容。仓库使用L<n>-cleared标签标记每个关卡通关时的完整状态快照，这些快照可直接用于下游智能体评估实验。从现有数据中还可推导出审议时间分析、失败原因分析和意外情况分析等维度。
The ARC-AGI-3 ls20 Agent Trajectories dataset records the full gameplay trajectories of the agy-gemini3.1pro model during its playthrough of the ARC-AGI-3 game ls20, and is part of the world model generalization experiments conducted by ARA. The core feature of this dataset is that the AI agent constructs a structured world model (dubbed Agent-Native Research Artifact) in real time during gameplay, and leverages this model to solve levels that cannot be cracked solely through initial exploration. The dataset adopts an independent storage structure for each harness×model×game combination, and data from different models playing the same game are stored in sibling repositories to facilitate comparative analysis. The data are synchronized in real time during gameplay, with updates pushed immediately each time the world model is finalized. The dataset includes records across multiple dimensions: the reasoning trace directory saves the world model construction process and logical reasoning trajectories of the agent; the action trace directory records the actions and decision-making basis of each sub-agent in JSON Lines format; the solution directory contains level progress snapshots and winning action scripts; the recording files provide frame-level gameplay recordings; additionally, it also includes rich content such as round-level logs, tool logs, API scorecards, world model prediction ledgers, and resource consumption statistics. The repository uses the L<n>-cleared tag to mark the complete state snapshot of each level upon clearance, and these snapshots can be directly used for downstream agent evaluation experiments. Additional derivable dimensions from the existing data include deliberation time analysis, failure cause analysis, and contingency analysis.
数据集概述
该数据集名为 ARC-AGI-3 ls20 Agent Trajectories (agy-gemini3.1pro)，记录了 agy-gemini3.1pro 模型在 ARC-AGI-3 游戏的 ls20 关卡中的完整游戏轨迹。它是 ARA-as-world-model 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（Agent-Native Research Artifact），并基于该模型攻克无法通过冷探索解决的关卡。
关键信息
- 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
- 数据集名称：ARC-AGI-3 ls20 Agent Trajectories (agy-gemini3.1pro)
数据布局
数据集包含以下文件及目录：
| 路径/文件 | 描述 |
|---|---|
reasoning/ara-ls20/ |
智能体的世界模型及推理轨迹（包含 logic/、trace/、staging/ 子目录） |
traces/ |
每个子智能体的动作-推理轨迹，每行一个动作及其原因（JSON 格式） |
solutions/GAME.md |
状态快照：进度、机制、注意事项、搁置原因 |
solutions/replay/ |
每个关卡的获胜动作脚本（文件名如 L<n>.txt） |
recordings/ |
按帧记录的日志，每会话一个文件（格式：<game>-<hash>.<agent>.<guid>.recording.jsonl） |
episodes/ |
按回合合并的运行日志 |
logs/ |
日志及自由格式会话笔记（git 历史保留之前会话） |
scorecards/ |
ARC API 评分卡（如保存） |
predictions.jsonl |
世界模型预测记录：置信度及后续验证结果 |
accounting/ |
本游戏的账本行（令牌/步骤）、指标部分、ARA 增长曲线 |
demo/ |
消融/展示工件（如存在） |
sessions_index.json |
每个记录会话的条目（智能体、GUID、帧数、最终状态） |
manifest.json |
来源信息：框架×模型 ID、智能体模型、仓库提交、协议哈希、同步时间 |
特殊标记
仓库标签 L<n>-cleared 标记每次关卡通关时的完整记录，作为下游智能体评估实验的 holdout-ready 快照。
衍生分析说明
决策时间（动作间间隔）、死亡分析及意外分析可从 recordings/ 和 traces/ 中推导得出，不单独存储。
