HAI HAI
Home ProjectsProfessional portfolio Contact

ARC-AGI-3 research

tabula-rasa

An experimental ARC-AGI-3 agent that explores unfamiliar interactive puzzles, with open experiment traces, a technical paper and reproduction instructions.

PythonLLM agentsARC-AGI-3
View source ↗Try demo
tabula-rasa project preview

Interactive preview

Try a demo

Explore a tiny grid. This is a small browser-only illustration, not the full application. AI responses are scripted; model and desktop features are not running here. Inputs stay on this page and reset when you reload.

Result

Ready. Change the sample and run the demo.

Explore the real workflow

Read the paper and repository setup.

The full app's capabilities and requirements are described below.

What it is for

Research on interactive reasoning agents, replay caching, action planning and measured evaluation.

Study how agents infer puzzle mechanics from observation and how planning choices affect action efficiency.

Inspect reasoning experiments
Study exploration, planning and replay from trace evidence.

Get started

Read the technical paper and reproduction guide before running the experiments.

View source ↗

Features and workflow

Interactive exploration
Discover game rules through observed outcomes.
Reproducible experiments
Inspect traces, technical paper and reproduction instructions.
  1. Read the paper and repository setup.
  2. Prepare the documented model and game environment.
  3. Run the reproduction instructions and inspect the resulting traces.

Architecture

The repository describes a persistent Python sandbox, game sessions, replay caches and scheduling across games. The technical paper documents a separate cascade-patch experiment; its evidence must be evaluated in its own scope.

Privacy and limitations

An experimental research project. Offline completion-ceiling measurements and competition scores use different scales and must not be compared directly. Inherited harnesses and methods are credited in the repository and paper.

Status

Public research code and paper; no hosted puzzle-solving service is offered. Results are reported experiments, not independently reproduced by this website.

Availability
Public source and technical paper
Status
Experimental research

Evidence

Technical paper and reported experiment evidence
Repository paper · reviewed 2026-10-22
Read the paper for inherited work, measured results and limitations.

PDF paper
Repository PDF · reviewed 2026-10-22
Exported research artifact.

Reproduction instructions
Repository instructions · reviewed 2026-10-22
Requires the documented environment; not executed by this website.

Questions

Are the scores interchangeable?

No. The repository distinguishes competition scores from offline completion-ceiling measurements.

Explore all projects →