Less is more: Meta study shows shorter reasoning improves AI accuracy by 34%

Meta researchers The User and Jerusalem Hebrew University In fact, they found that there was less to “think” to “think” of large language models to improve their performance on majority thinking tasks.

This learn Today, in the AI ​​systems released, more short reasoning processes have caused more accurate results in case of significantly reducing the calculation costs.

“In this case, we protest against the hypothesis that shows the results of long thinking chains in the ability to think of better thinking,” Write authors in their paper “Don’t overdo it. Prefer to shorter thinking chains for improved llm thinking“

Research is contrary to the prevailing Trend in the development of the AI, which allows companies to expand the sources of computing to allow for the length of the models, “Thinking chains“- Step-by-step trajectories used by AI systems to solve complex problems.

AI accuracy, models jump 34% when using shorter thinking chains

Researchers are in the same way of thinking, “the more likely chains are more likely to answer correctly – the same question is about 34.5% for the same question.” This finding is true among many leading AI models and criteria.

“When showing effective results, [extensive reasoning] There is a significant calculation costs and the results of the result, “the authors who point to how these systems are currently being placed.

According to these findings, a team called a novel named “Short m @ k“In parallel, perform many reasoning attempts, but stops the calculation after the first few processes are completed. The last answer is selected by voting voting between shorter chains.

The new ‘short m @ k’ method reduces calculation costs 40% while increasing performance

For organizations that placed large AI justification systems, the results may be important. Researchers could reduce the methods of up to 40% while keeping the level of performance as standard approaches.

“Short-3 @ K, while short-1 @ k, still sounds between all computing budgets, still faster (reduction up to 33% wall times),” paper states.

Michael Hassid, the lead author and team of the paper also improved the performance of AI models in more convening samples – a more fundamental assumption in the development of AI.

“Training about short ones causes better performance,” researchers write. “On the contrary, Fineruning in S1 increases justification time without an important performance profit.”

Tech giants can save millions by carrying out the approach of “end”

Findings come in a critical time for the AI ​​industry, because the companies are increasingly powerful models that consume very large sources of computing.

“Our Croples are allowing our frustration to think that the reasoning ways are not translate to the adjustable performance and emphasize that it can lead to unknown results.

‘This study stands as opposed to other prominent approaches. Previous prestigious studies, including Openai’s work “Thoughtful” desire and “self-roundness“Methods have generally defended for broader justification processes. This also builds in recent work as Princeton and Google Deepmind”Thoughtful“Frame and Carnegie Mellon”Deny self“Methodology examining different approaches to AI thinking.

Research evaluating AI investments for technical decision makers, the larger and more computing intensity indicates that the intensive is not always better. Research is not a raw calculation power, not optimizing efficiency, optimizing the potential costs and performance improvements.

Using a scale, it does not only keep the power of the EU, the EU does not only keep the power – it makes machines even smarter. Sometimes, even the artificial intelligence from old wisdom benefits: Do not exceed it.