Google’s Flan AI Makes Language Models Smarter Without More Data

Researchers improved AI models by “instruction finetuning” them on 1,800+ tasks and adding chain-of-thought reasoning data. The result: Flan-PaLM significantly outperforms the original PaLM across benchmarks, improves reasoning, works better in zero-shot settings, and delivers stronger real-world responses—all with minimal extra compute.