ARCTraj: Human Reasoning Trajectories for ARC-AGI

Dataset Summary

Standard ARC datasets provide only input–output grid pairs.
ARCTraj provides the missing reasoning process: fine-grained human action trajectories collected during the ARC-AGI task-solving.

ARCTraj enables process-level supervision, imitation learning, trajectory modeling, and reinforcement learning on ARC-style environments.

---

TL;DR

• ~10,000 human trajectories
• 400 ARC-AGI-1 training tasks
• Object-level action sequences
• Timestamps and success labels
• Collected via O2ARC 3.0 (IJCAI 2024 Demo)
• Compatible with ARCLE and JaxARC-style environments

---

Dataset Details

• Curated by: Data Science Lab (DSLab), GIST
• License: CC-BY-NC-4.0
• Data format: CSV interaction logs
• Language: Not applicable (interaction logs)

ARCTraj contains anonymized human interaction logs recorded during ARC task-solving sessions.

Each trajectory records the full sequence of intermediate reasoning steps rather than only the final output grid.

---

Dataset Structure

The dataset is provided as interaction logs (ARCTraj.csv).

Each row corresponds to a problem-solving session and contains:

Column	Description
logId	Unique log session ID
userId	Anonymized user identifier
taskId	ARC task identifier
actionSequence	Ordered list of object-level actions performed by the user
startedAt	Session start timestamp
endedAt	Session end timestamp

From these logs, trajectories can be reconstructed into:

• State sequences
• Action sequences
• Timestamps
• Success labels

These representations enable MDP-compatible preprocessing for RL-style environments.

---

Quick Start

from datasets import load_dataset

dataset = load_dataset("SejinKimm/ARCTraj")

sample = dataset["train"][0]

print(sample.keys())
print(sample["taskId"])
print(sample["actionSequence"][:5])

---

Intended Uses

ARCTraj can be used for:

• Imitation learning
• Offline reinforcement learning
• Learning from demonstration (LfD)
• Decision Transformer training
• Process supervision for ARC reasoning
• Human-initialized policy learning

---

Research Using ARCTraj

Subsets of ARCTraj have been used in:

• PPO training with ARCLE-style preprocessing:
  https://github.com/GIST-DSLab/PPO_Solve

• Latent diffusion–based learning on ARC tasks:
  https://github.com/GIST-DSLab/LDCQ

• Decision Transformer–based trajectory modeling:
  https://github.com/GIST-DSLab/IntentionLearning

• Grid and trajectory augmentation (SOLAR-Generator):
  https://github.com/GIST-DSLab/SOLAR-Generator

---

Data Collection

Trajectories were collected via O2ARC 3.0 (IJCAI 2024 Demo):
https://o2arc.com

Participants solved ARC-AGI tasks through an interactive object-level interface.
All intermediate interactions were logged and anonymized.

---

Risks and Limitations

• All user identifiers are anonymized.
• No personally identifiable information is included.
• The dataset is intended strictly for research use.
• Commercial use is restricted under CC-BY-NC-4.0.

ARCTraj reflects the distribution of participating users and may contain biases in expertise and strategy diversity.
Undo/Redo operations and selection-operation separation may require preprocessing for certain RL environments.

---

Citation

If you use ARCTraj in your research, please cite:

@inproceedings{kim2026arctraj,
  title={ARCTraj: A Dataset and Benchmark of Human Reasoning Trajectories for Abstract Problem Solving},
  author={Kim, Sejin and Choi, Hayan and Lee, Seokki and Kim, Sundong},
  booktitle={KDD Datasets and Benchmarks},
  year={2026}
}

---

Links

• Paper: https://arxiv.org/abs/2511.11079
• Project page: https://arctraj.github.io/
• O2ARC 3.0: https://o2arc.com

---

ARCTraj bridges symbolic ARC reasoning and learning-based approaches by providing access to human problem-solving processes.