Harvard-linked study finds reasoning models get stuck on their first few words

A Harvard-linked study found that AI reasoning models tend to get stuck on their first few words, limiting their ability to follow instructions. The research suggests simple interventions could improve instruction-following capabilities in these models. Better reasoning performance could reduce AI development costs by making models more efficient. The findings matter for crypto AI protocols and tokenized AI projects that depend on model performance and cost efficiency. Watch for follow-up research and adoption of these interventions by AI developers and decentralized AI networks.