Home

Awesome

Language-complete Abstraction and Reasoning Corpus (LARC)

This repository contains the LARC dataset and supporting assets

The entire dataset can be browsed at the explorer interface take a look!

Here's a quick 5 minutes slideslive video explaining this work

"How can we build intelligent systems that achieve human-level performance on challenging and structured domains (like ARC), with or without additional human guidance? We posit the answer may be found in studying natural programs - instructions humans give to each other to communicate how to solve a task. Like a computer program, these instructions can be reliably "executed" by others to produce intended outputs."

A comprehensive view of this dataset and its goals can be found in Communicating Natural Programs to Humans and Machines (Neurips Dataset and Benchmark, 2022)

LARC is curated from a communication game, where one participant, the describer solves an ARC task and describes the solution to a different participant, the builder, who must solve the task on the new input using the description alone. The successful descriptions are "language-complete" in a sense that it fully captures the underlying ARC task in the absence of the original input-output examples.

<p align="center"> <img src="https://raw.githubusercontent.com/samacqua/LARC/main/assets/collection.jpg" alt="drawing" width="75%"/> </p>

Citation

@article{acquaviva2021communicating,
  title={Communicating Natural Programs to Humans and Machines},
  author={Acquaviva, Samuel and Pu, Yewen and Kryven, Marta and Wong, Catherine and Ecanow, Gabrielle E and Nye, Maxwell and Sechopoulos, Theodoros and Tessler, Michael Henry and Tenenbaum, Joshua B},
  journal={arXiv preprint arXiv:2106.07824},
  year={2021}
}

The original ARC data can be found here The Abstraction and Reasoning Corpus

Contents

language-guided program synthesis code can be found here

GPT4 (vision only) program induction results can be found here

License

The dataset is licensed under the Creative Commons Attribution 4.0 International License

All supporting code follows the MIT License