ARC Prize: A Guide to DSL, LLM-Guided and Test-time Training Approaches

This comprehensive tutorial explores approaches to solving the Abstraction and Reasoning Corpus (ARC) Prize challenge, providing detailed explanations of Domain Specific Language (DSL), Large Language Model (LLM)-guided, and test-time training methodologies. Learn how to solve ARC problems manually before diving into three primary solution approaches with practical demonstrations. Discover how to prepare ARC training and evaluation data, implement a detailed DSL approach, utilize LLM-guided search techniques, and apply test-time training with depth-first search. Access the open-source repository for hands-on practice and explore opportunities to join or sponsor the Trelis ARC AGI Team. Perfect for AI researchers and enthusiasts interested in abstract reasoning challenges and cutting-edge AI problem-solving techniques.

Syllabus

0:00 What is ARC AGI?
1:00 Solving ARC Manually
3:59 Broad approaches to solving ARC AGI
8:04 Simple DSL Approach Domain Specific Language
12:33 Simple LLM Guided Program Approach
16:04 Simple Neural Net Approach with beam and depth-first search
28:24 Overview of ARC training and evaluation data and preparation
33:44 Detailed DSL Approach Domain Specific Language
48:47 LLM Guided Search in detail
1:09:00 Test time training approach, with depth-first search
2:00:00 Apply for the Trelis ARC AGI 2 team