The researchers find that you do not need a lot of data to train LLMS on thinking tasks

LLMS models can learn complex thinking tasks without relying on large data collections, according to New study By researchers at the University of Shanghai Jiao Tong. The results of their findings show that through a small set of well -coordinated examples, you can train LLM on tasks that he believed was required tens of thousands of training counterparts.

This efficiency is due to the underlying knowledge that modern LLMS gets during the pre -training stage. Although new training methods become more efficient in data and their account, institutions may be able to create dedicated models without the need to reach the resources of large artificial intelligence laboratories.

Less is more (Limo)

In their study, researchers challenge the assumption that you need large quantities of data to train LLMS on thinking tasks. They offer a concept “less is more” (Limo). Their work works on his head Previous research Which showed that LLMS can be aligned with human preferences with some examples.

In their experiences, they have proven that they could create a Limo data collection for complex sports thinking tasks with a few hundred training examples. LLM, which was seized on the data collection, was able to create a complex A series of ideas (COT) thinking chains that enabled them to accomplish tasks at a very high success rate.

For example, a QWEN2.5-32B-Instruct The models seized on 817 training examples were chosen based on 57.1 % of the accuracy in the very difficult AIME standard and 94.8 % over mathematics, the performance surpassed the models that were trained on a hundred again. He also recorded higher on standards of thinking models such as QwQ-32B-PREVIEW (A copy of the QWEN model was trained to think) and Openai O1-PREVIEWBoth were trained with larger data and resource account.

Moreover, circulating the forms trained on Limo to the examples is very different from their training data. For example, on Olympiadbench The scientific standard, the LIMO model is outperforming the QWQ-32B-PREVIEW, and on the challenge GPQA Standard66.7 %, near the leading result in Openai-O1-PREVIEW by 73.3 %.

What does it mean to the institution AI?

LLMS customization is an attractive state of use of the institution’s applications. Thanks to techniques such as A generation for retrieval (Flaq) and Learning within the contextLLMS can be customized to use detailed data or perform new tasks without having a good expensive adjustment.

However, thinking tasks often require training and LLMS control. The belief was widely that such tasks required large quantities of training examples with thinking chains and very detailed solutions. The creation of these data collections is slow and non -practical for many applications and companies.

Recently, the researchers showed this Purifying learning approach Models can be enabled to train themselves on thinking tasks by generating many solutions and choosing those that work better. While this approach requires less manual effort, it still requires expensive arithmetic resources far from the reach of many institutions.

On the other hand, the formulation of a few hundred examples is the endeavor that many companies can address, which makes specialized thinking models within reach a wide range of organizations.

“This discovery has profound effects on artificial intelligence research: it indicates that even the complex thinking capabilities at the level of competition can be effectively devised through minimum but coordinated training samples.”

Why Limo works

In their experiences, researchers define two main reasons behind LLMS learning complex thinking tasks with less examples.

First, modern basic models have been trained in a very large amount of Sports content and symbol During training. This means that these LLMS already have a rich logical knowledge in its criteria that can be activated through carefully made examples.

Second, post -training techniques showed that allowing models to generate expanded thinking chains greatly improves their ability to think. In essence, the forms are given more time to “think” to empty and apply their prior knowledge more effectively.

“We assume that successful logic highlights the synergy of these two factors: the rich knowledge that was previously trained and sufficient arithmetic resources at the time of reasoning,” researchers write. “These developments collectively indicate an amazing possibility: If the models have rich logical knowledge and are given sufficient mathematical space, then stimulating their thinking capabilities may only require a small number of high -quality training samples that encourage extended deliberations, instead of huge click data sets .

According to the results of the researchers, the creation of useful Limo data groups depends on choosing appropriate problems and solutions. Data coordinators should give priority to difficult problems that require complex thinking chains, various thinking processes and integration of knowledge. Problems must also deviate from the distribution of the model training to encourage new thinking methods and force it to generalize.

Accordingly, the solutions should be clearly organized and well organized, while adapting the steps to think with the complexity of the problem. High -quality solutions should also provide strategic educational support by gradually building understanding through carefully organized interpretations.

“By focusing on a group of lower logical thinking chains, but, we embody the basic principle of the Calcium: high -quality demonstrations, instead of the massive data, is the key to opening complex thinking capabilities,” the researchers wrote.

The researchers have Introduce the symbol and data Used to train Limo models in their experiments. In the future, they plan to expand the concept to other fields and applications.

Daily visions about business use cases with VB daily

If you want to persuade your boss at work, you have covered VB Daily. We give you the internal journalistic precedence over what companies do with obstetric artificial intelligence, from organizational transformations to practical publishing operations, so that you can share visions of the maximum return on investment.

Read with us privacy policy

Thanks for subscribing. Check more VB news bulletins here.

An error occurred.