This too much thought can be a problem for the world:

Large language models such as Openai O1 (LLMs) represent significant progress in performance, but are still far from ideal efficiency. Researchers at the University of Tencent AI Lab and Shanghai Jiao Tong have published a comprehensive study on a repeating problem in these models that are trying to imitate human reasoning: over thinking (or “excessive reflection” in Portuguese).

The study analyzes the excessive use of coins, high -computing power consumption and waste of resources in the processing of productive models such as similar Chinese rival O1 and Deepseek R1. These Artificial intelligence tries to simulate the human cognitive process The chain of thought (to formulate answers in an approach known as “chain of thought in COT or Portuguese. However, this method can consume more powerful – more powerful – up to 1.953% without need.

In the article, researchers show the phenomenon with a simple question: 2 + 3 how much? The test was applied to several popular LLMs such as GPT-4O, Gemini Pro and Claude-3.5 and compared with the QWQ-32B prevention, a rational model of the QWEN team.

The results show a striking contrast:

• Traditional LLMs provided the correct answers that consumed less than 10 coins (except QWEN2.5-MATH-72B using almost 50 coins);
• The QWQ-32B measure used 901 coins to answer the same question.

The rational model of the QWEN team detailed the 10 different solutions that all reached the same result: 2 + 3 = 5.

Although the question is simple, QWQ-32B Prevention cannot separate it from more complex inputs. Therefore, the first solution reflects the problem until there is no other alternative, even if the first solution is already correct.

“In the figure, we observed that the first tour of the first solution has already the right answer. Subsequent solutions that make up most of the coins produced, Do not increase accuracy”, Emphasizes the work.

After testing, the researchers found that IAS received the right answer in the first attempt in 92% of the scenarios. The problem of over Thinking was more repeated in simpler mathematical issues.

Why is there a problem?

Productive artificial intelligence requires significant calculation power to stay active. Increasing the need for processing directly affects energy consumption and the use of components based on data centers. This is a special problem for companies with central platforms such as Openai, Deepseek and Google, To meet the increasing request of the users, you need to expand their servers continuously..

For the user The biggest problem is the consumption of the context window. Thought chain technique uses more coins than usual, and these coins describe the context window – for example the area where the problem is added. Although it does not make a big difference in these simple demands, It can significantly affect more complex requests.

Excessive thinking is just a problem when it is in vain

However, the application of COT technique to advanced language models represents an important progress. The ability of documenting this reasoning line while producing answers It is extremely useful for artificial intelligence education and development of soulsArtificial Intelligence Researcher Billy Garcia and Billy Garcia, the founding partner of Abstrakt Gen-Ai Tecmundo.

“Basically, the loss of productivity of 1.953% takes place only when the model is selected inappropriate,” Garcia explains. “However, it is necessary to access this chain of reasoning, especially for certain use of the research.”

“Then, Users should not resort to advanced models to answer insignificant questions“How much 2 + 3?” He points to the expert.

Is there a solution for AI Overtying?

The article is investigating different strategies to reduce excessive reflection and make rational models more efficient. That Solutions include educational methods optimizedIt is like self -training that uses the data set samples created by AI to educate and develop AI itself.

“Self -trade takes place through different treatment methods and aims to make the model more efficient without compromising accuracy for more complex tasks.” Basically, It’s like preparing the model to “think less”.

Among the recommended approaches:

• Supervised Treatment: Improvement of Models based on positive synthetic data;
• Direct Preference Optimization: Model training, considering the preferred response by people;
• Reasoning Preference Optimization: Adding negative reasoning records to avoid unnecessary repetitions;
• Simple Preferably Optimization: Fine adjustment to align a reward function to respond production metropy.

However, alone, These solutions do not completely eliminate the exaggeration. The study said, “Shorter sample answers increase the efficiency of O1 type models, but they still suffer from excessive reflection.”

Therefore, the article proposes complementary methods to determine when AI received the correct answer, for example:

• The first correct solution (FCS): The first response defines the right one;
• FCS + Reflection: Allows AI to reflect the accuracy of the first answer, and provide a more reliable result in the second attempt;
• Enthusiastically, various solutions: If previous outputs are not consistent, it adds new reflection strategies.

By combining these strategies, researchers He observed a significant decrease in coin consumption And in computational demand, without compromising the cognitive capacity of rational artificial intelligence compared to traditional LLMs.

Does everyone need a rational artificial intelligence?

O1 and Deepseek-R1 represents remarkable progress in the development of productive artificial intelligence, but Applications are not about daily use. “The complexity of these models, scientific research or complex decision -making, such as deep reasoning problems that require deep reasoning, he explains.

In daily life, in tasks such as simple software development, review of short texts and other insignificant applications, It is likely to consume unnecessary coins.

Currently, Openai O1 is available with “limited access ında in the Chatgpt Plus subscription. During the preview phase, the Model offered a context window up to 128,000 to the coin, distributed between O1 prevention (32 thousand coins) and O1-Mini (65 thousand coins) versions.

Thus, while applying to artificial intelligence aid, It is important to choose which model to use well.. This can help recover not only faster for answers, but also to recover the coins that may be useful in a later consultation.