The researchers find you don’t need data data to train LLMS in the discussion

Languages ​​(LLMS) can learn the conversation without dependence on the main data, according to a New Study With Shanghai Jiao Ton Ton Ton. The results indicates that it is only a small group, you can teach llm in the work of being considered thousands of thousands.

This success is caused by information that Llms finds on the training session. With new teaching techniques, a number of businesses, businesses, businesses can make unchanged varieties without using great Ai Labs.

The minimum one is more (limo)

In study, researchers are opposed to the idea you need more data to teach LLMS in the discussion. Causes the idea of ​​”less and” (limo). Their job builds on top of Ancient research This showed llms may be equal to what they like to love with a few examples.

In their efforts, he showed that they could make the Data Data of Limo’s difficulty in number of hours of training. A good photographer on Dataset was able to make a problem Principle-emotional .

For example, a Qwen2.5-32b-consultants A Well-Trained Picture of 817 Examples of Teaching Training For 57.1% on Math ATIMRA HADNY RIGHT, SPECIAL TRANSMENTS. Was the most thoroughly on Benchmarks than equal types as QWQ-32B-PANTENILO (the type of number of qwen version of the data discussion) is Tsetai O1-noiseAll who were trained by large and compound equipment.

In addition, the best examples of limo differs from what teaches. For example, at the Olympadgench Science sign, the form of gentle limo QWQ-32b-see, and in difficult situations GPQA BenchmarkFounded 66.7% correctly, near the 73.3% open field.

What does it mean for Ai business?

LLMS change is very important to use business. Due to the skills such as the generation of restoration (Room) and detailed studyLlm can be used to use use or perform new jobs without the cost of cost.

However, the idea of ​​arguing often requires education with LLM. The actual beliefs have been used if this requires great examples of training and detailed charts and answers. Making money if this is running out and unchanged on a job with many companies.

Recently, investigators have shown that Pure learning learning is close It can allow varieties to train work in the production of many answers and select what works better. Although the process requires a minimum effort, it still requires the expensive things that are impossible in many businesses.

On the other hand, scatting a hundred examples and efforts that many companies can handle corresponding variables to multiple corporations with multiple corporations.

“This discovery has a great meaning in the wise investigation: It indicates that even the difficult skills of competition do well but the teaching,” investigators wrote.

Just one reason

In their trials, researchers find two major reasons that LLMs can learn the conversations with a few examples.

First, the first part of the state of the state has been very well trained Maths and number During the training time. This means that llms have enough information on their starting parts that may occur with modest examples.

Second, new training techniques have shown that allowing colors to make the best of the viewers. In this way, giving examples of time, “thinking” allows them to produce and use their knowledge well-trained.

“We’ll reap what to do for a good deal from the Syrnergy of the two elements: the heavy knowledge and fully trained and helpful enough for time,” researchers wrote. This makes it possible to be done: Paper. “

According to investigating research, I’m making useful UPSOSETS when choosing the correct problems and answers. The Data Coins should make an important challenge that requires a difficult chain, a variety of ways they think and get to know each other. The problems must also be separated and partitioning of the species to encourage new reasons to approach and force them to despair.

So, the solution should be clear and well-organized, and thoughts followed by the problem of the problem. The highest responses should also provide training effective behaviors using slowly understanding well-being.

“Reflecting on the best of the choices, we have a bright focus of Limo: high-quality shows, instead of opening the formation of the word,” investigated.

The researcher has He pulled out the number with data used to teach the species of limo in an attempt. In the future, prepares to build up the idea of ​​other areas and software.