Luca Favole
Co-Evolving Representation and Program: A Solver for ARC-AGI.
Rel. Giovanni Squillero, Giorgio Mongardi, Alberto Paolo Tonda. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (3MB) | Preview |
Abstract
Machines write complex code and solve olympiad problems, yet fail at puzzles most people solve without training. The Abstraction and Reasoning Corpus (ARC) makes that gap measurable. Each task gives a few coloured grids and asks for the output grid a further input should produce. Because every task is new, no training distribution can be fitted and data contamination cannot inflate a score. Humans average above 70% on the training set, while plain language models manage about 10%; reasoning models reach above 90%, but on subsets of the tasks and at dollars per puzzle. An alternative symbolic route would be to search for a short program fitting the examples, which moves the difficulty to the representation: over raw pixels the programs needed are hopelessly long.
The question this thesis takes up is whether that representation can itself be a large parameterized space, searched per task alongside the program and within a budget that fits one machine
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
