New Transformer Model Achieves 44% Score on ARC-AGI-1 Benchmark

Sep 1, 2026

AI Summary

A newly developed transformer model has achieved a score of 44% on the ARC-AGI-1 benchmark, surpassing previous models in speed and efficiency. The model is open source and aims to improve sample efficiency in AI research while reducing costs for experimentation.

The new transformer model was trained in 1.5 hours on a 5090 GPU and outperforms many existing large language models (LLMs). It also scored 7% on the ARC-2 benchmark.
This model is an upgrade from a previous version, designed to be faster, better, and cheaper while remaining open source.
Key changes include training only on output tokens, increasing training data by incorporating non-overlapping tasks from ARC-2, and avoiding data leakage by filtering repeated puzzles.
The model's performance improvements are attributed to better representations and architectural changes, while costs can potentially be reduced significantly with optimized GPU code.
The author emphasizes the importance of sample efficiency in AI and plans to continue exploring new research ideas to push the limits of current deep learning methods.
The ARC benchmark has been a focal point for testing AI capabilities, and the author argues that it remains relevant for assessing sample efficiency and meta-learning.
There has been significant discussion and debate among researchers regarding the implications of the results, with some criticisms of existing models and methodologies in the field.

arc-agi-1researchartificial intelligencemachine learninginnovation

New Transformer Model Achieves 44% Score on ARC-AGI-1 Benchmark

Related Stories

AI Models Successfully Decode Long-Unsolved Enigma Messages

OpenAI agents reportedly attempt to access secure databases and government websites

Anthropic AI Claims Discovery of Enzyme System Similar to CRISPR Technology

Anthropic's biology lab claims significant enzyme discovery using AI