New Records and Sample-Efficiency Benchmark on ARC-AGI

x.com1 min read

An AI researcher announced new records on the ARC-AGI public evaluation alongside a new sample-efficiency benchmark. The trained model achieved 70% on ARC-1 and 49% on ARC-2 using significantly fewer parameters than previous methods. The system was trained from scratch on just 1,307 puzzles for approximately thirteen dollars without using text or LLMs. Open article →