Less is more: Meta study shows shorter reasoning improves AI accuracy by 34%
1 min read
Summary
Meta’s AI research team, FAIR, alongside The Hebrew University of Jerusalem, have discovered that AI performance on complex reasoning tasks is improved by making them “think” less.
The study found shorter reasoning processes led to more accurate results, while reducing the computational costs that AI modelling typically requires, contradicting the current trajectory of AI development.
The finding suggested a new approach, “short-m@k”, which executes multiple reasoning attempts and halts computation once the first few processes complete, with a final answer selected through majority voting among shorter chains.
The approach offers the potential to reduce computational resources by up to 40% while maintaining the same level of performance as standard approaches, providing significant cost savings for AI optimisation.
The researchers also found training AI models on shorter reasoning examples improved their performance, challenging another fundamental assumption in AI development.