Less is more: Meta study shows that shorter reasoning improves the AI ​​-nuclearity by 34%

Researchers from Meta’s Fair team And The Hebrew University of Jerusalem have discovered that forcing large language models to ‘think’ less, actually improves their performance in complex reasoning tasks.

The study Today it appeared that shorter reasoning processes in AI systems lead to more accurate results, while the calculation costs are considerably reduced.

“In this work we dispute the assumption that long thinking chains result in better reasoning options,” write the authors in their paper entitled “Don’t think about it. Prefer shorter thinking chains for improved LLM reasoning. “

The research is in contradiction with the prevailing trend in AI development, where companies have invested heavily in scaling up computer sources to enable models to carry out extensive reasoning by long -term “thinking chains“Detailed step-by-step processes that use AI systems to solve complex problems.

AI -nouvulity stands 34% when models use shorter reasoning chains

The researchers discovered that within the same reasoning task “shorter reasoning chains are considerably more likely to give the correct answers – up to 34.5% more accurately than the longest chain that was sampled for the same question.” This finding kept where in several leading AI models and benchmarks.

“While you show impressive results, [extensive reasoning] Brings significant calculation costs and inference time, “the authors notice, pointing to a substantial inefficiency in how these systems are currently being used.

Based on these findings, the team developed a new approach called ‘Short-M@K“Implement those multiple reasoning attempts in parallel, but the calculation stops once the first few processes are completed. The final answer is then selected by the majority of voting between these shorter chains.

New ‘Short-M@K’ method drops the computer costs by 40% and stimulates performance

For organizations that use large AI reasoning systems, the implications can be considerable. The researchers discovered that their method could reduce the calculation sources by a maximum of 40%, while the same performance level as standard approaches could be maintained.

“Kort-3@K, although slightly less efficient than short-1@K, consistently surpasses majority votes over all calculation budgets, while it is still considerably faster (up to 33% wall-time reduction),” the paper says.

Michael Hassid, the main author of the newspaper, and his team also discovered that training AI models about shorter reasoning examples improved their performance -challenging another fundamental assumption in AI development.

“Training on the shorter leads leads to better performance,” the researchers write. “Conversely, the fine tuning on S1-Long increases the reasoning time without significant performance profits.”

Technical giants can save millions by implementing the approach of “Don’t Overhink it”

The findings come in a critical time for the AI ​​industry, because companies are racing to use more and more powerful models that use huge areas of calculation.

“Our findings suggest that reconsidering the current methods of testing time that are calculated in reasoning in reasoning LLMs, and emphasizes that longer ‘thinking’ cannot necessarily translate into improved performance and contraintifically lead to broken results,” the researchers conclude.

“This research is in contrast to other prominent approaches. Previous influential studies, including the work of OpenAi on “Debit” instructions and “self -consistency“Methods have generally argued for more extensive reasoning processes. It also builds on recent work such as Princeton and Google DeepMind’s”Toom of thoughts“Framework and Carnegie Mellon’s”Self -reflecting“Methodology, which have investigated various approaches of AI reasoning.

For technical decision makers who evaluate AI investments, the research suggests that larger and more computational intensive is no better. The study points to potential cost savings and performance improvements by optimizing for efficiency instead of rough computing power.

In an industry obsessed by scaling up, it appears that learning AI is more concise, not only keeps the computing power – it also makes the machines smarter. Sometimes even artificial intelligence of the age -old wisdom benefits: don’t think about it.