hellohawaii/ARC3-human-think-aloud

ARC3 Human Think-Aloud

ARC3 Human Think-Aloud records human learning and problem solving in 25 public ARC-AGI-3 games. It pairs gameplay actions and visual observations with concurrent think-aloud recordings and time-stamped transcripts, making it possible to study how understanding develops over the course of an interaction. The dataset also includes LLM agent trajectories on the same games for comparative analysis.

The dataset accompanies How Humans Learn Unfamiliar Worlds: A Think-Aloud Dataset on ARC-AGI-3, by Shuzhi Liu, Archana Warrier, Dat Nguyen, Chengqi Zheng, Kevin Ellis, and Yewen Pu.

Dataset Overview

	Human participants	LLM agents
Games	25	The same 25 games
Participants / configurations	64 participants	10 model–harness configurations
Trajectories	144 successful trajectories	250 trajectories, including successful and unsuccessful runs
Actions	136,876	99,930
Verbal records	Think-aloud audio and transcripts	Available public reasoning records

Human recordings are provided as 147 game/session audio clips. The shared annotation framework covers 354 game mechanisms and three aspects of interaction: mechanism triggering and recognition, action purposes, and errors and recovery.

The data support research on world-model learning, exploration, hypothesis testing, planning, and error recovery, as well as comparisons between human and LLM approaches to interactive problem solving.

Data Collection

Human participants were recruited through Prolific. After study instructions, consent, a microphone check, interface training, and a think-aloud attention check, participants explored unfamiliar games while verbalizing their thoughts. Gameplay actions and speech were recorded concurrently. Audio is accompanied by automatically generated transcripts.

The agent collection contains 100 trajectories from direct LLM configurations and 150 from configurations using VISTA, Schema, OPINE-World, or the Provider Adapter. Each configuration contributes one trajectory for each of the 25 games, with its actions, outcome, and available public reasoning.

Annotation Framework

Annotations connect behavior and verbal evidence to three questions:

Annotation	What it captures
Mechanism triggering and recognition	When a game mechanism is encountered and when the learner shows evidence of understanding it.
Action purposes	Whether actions pursue a goal, gather information, or serve both purposes, including exploration and hypothesis testing.
Errors and recovery	Mistaken mechanism beliefs, planning failures, execution errors, and subsequent corrections.

Game mechanisms were defined by the authors. Trajectory annotation combines programmatic trigger detection with AI-assisted analysis of actions and verbal evidence, with author spot checks. Each game's mechanisms.json defines its mechanisms; each trajectory's annotations.json contains the associated labels, action ranges, and supporting evidence.

Data Organization

Human data are organized by game / participant, and agent data by game / harness / configuration:

games/
└── ar25/
    ├── mechanisms.json
    ├── P02/
    │   ├── metadata.json
    │   ├── audio.mp3
    │   ├── actions.json
    │   ├── observed_actions.json
    │   ├── transcript.json
    │   ├── transcript.txt
    │   ├── annotations.json
    │   ├── timeline.json
    │   └── timeline.json.gz
    ├── P03/
    │   └── ...
    └── llm/
        └── direct/
            └── anthropic-claude-opus-5-high/
                ├── metadata.json
                ├── actions.json
                ├── reasoning.json
                └── annotations.json

Participants are identified by pseudonyms P01–P64, which are consistent across games. Human trajectory IDs combine the pseudonym and game, for example P28__ls20. Original recruitment identifiers are not included. A trajectory with multiple sessions has numbered audio, transcript, and timeline files, such as audio_01.mp3 and audio_02.mp3, in the same participant folder.

File	Contents
metadata.json	Trajectory information, session descriptions, and paths to the associated files.
audio.mp3	Think-aloud recording for the game/session.
actions.json	Canonical action sequence and trajectory information.
observed_actions.json	Logged human inputs, including undo operations and available timing information.
transcript.json / transcript.txt	Time-stamped utterances and a readable text version.
annotations.json	Mechanism, action-purpose, and error/recovery annotations.
timeline.json / timeline.json.gz	Game frames and their time-stamped links to actions; both files contain the same data.
reasoning.json	Public agent reasoning linked to action steps.

At the repository root, index.json lists games and trajectories, and metadata.jsonl provides a tabular index for the Hugging Face Dataset Viewer. The corpus split contains one index row per trajectory; audio and detailed records are available through the listed paths.

Transcripts use seconds relative to each session segment; timeline timestamps use milliseconds. Session metadata supplies the audio offset. Annotation action ranges are one-based and inclusive, referring to the canonical action sequence rather than positions in the observed-input log.

Getting Started

Install the Hugging Face client:

pip install -U huggingface_hub

Download and read one participant's data for a game:

import json
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="hellohawaii/ARC3-human-think-aloud",
    repo_type="dataset",
    allow_patterns=["games/ar25/P02/*"],
))
participant = root / "games/ar25/P02"

metadata = json.loads((participant / "metadata.json").read_text())
actions = json.loads((participant / metadata["actions"]).read_text())["actions"]
annotations = json.loads((participant / metadata["annotations"]).read_text())

segment = metadata["segments"][0]
audio_path = participant / segment["audio"]["path"]
transcript = json.loads((participant / segment["transcript"]).read_text())

print(f"Actions: {len(actions)}")
print(f"Audio: {audio_path}")
print(transcript["utterances"][0]["text"])

Remove allow_patterns to download the full dataset. To browse a particular game, use a pattern such as games/ar25/**.

Citation

If you use this dataset, please cite:

@misc{liu2026arc3humanthinkaloud,
  title        = {How Humans Learn Unfamiliar Worlds: A Think-Aloud Dataset on ARC-AGI-3},
  author       = {Liu, Shuzhi and Warrier, Archana and Nguyen, Dat and Zheng, Chengqi and Ellis, Kevin and Pu, Yewen},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/hellohawaii/ARC3-human-think-aloud}
}

License

The dataset is released under CC BY 4.0. Third-party game assets and agent source materials remain subject to their respective licenses.