Skip to content

tanchongmin/ARC-Challenge

Repository files navigation

ARC-Challenge

Refined Update (8 Oct):

Paper Uploaded at: https://arxiv.org/abs/2310.05146

Solves 50 out of 111 training set problems! No change to earlier 22 Jul update. Added a more refined UI to go through the .html files for the 111 problems. Jupyter Notebook included inside the folder "LLM_Expert_Agents_For_ARC_08Oct2023"

Based on the latest paper, intending to add more functions and to refine the agents to improve performance, stay tuned!

22 Jul:

Solves 50 out of 111 training set problems! Uses object and pixel abstraction spaces and helper functions to ground input-output relation!

  • Added ARC_Challenge_22Jul2023.ipynb : Latest iteration of GPT4 API to automate solving ARC Challenge - Note it can get expensive, running one iterative feedback loop cycle for one task costs about 30-40 cents.
  • Added ARC_Training_220723 folder to showcase results from this Jupyter Notebook on 111 training set problems that have <3k context length. This number caters for the iterative feedback loop additional information, and is likely to not exceed 8k maximum token length. 50/111 Solved

Overview

This contains a series of Jupyter notebooks (with date in the file name), to document the progress I have made at getting GPT4 to solve ARC.

Papers:

  • https://arxiv.org/abs/2306.03553 - An Approach to Solving the Abstraction and Reasoning Corpus (ARC) Challenge: My Lab42 ARC Essay Challenge submission, based primarily on my initial experiments documented in arc_challenge_basic.ipynb
  • I have more progress so far, and will write a follow-up paper shortly, after I finish experimenting on the ARC training set. The new approach involves:
    • Full end-to-end pipeline without human intervention
    • Sample input/output pairs to language description
    • Language description to list of functions via function grounding in human priors
    • Conditional function execution using if statements to check on conditions (I call this the Instructions Code Format)
    • Better processing of input grid via objects
    • Reflection-like pathway using environment feedback multiple times as in Voyager / Ghost in the Minecraft
    • Broad to specific grounding via efficient prompting

Videos:

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors