New technique helps the help

Thinking through chain thought (COT) – The management of the believers breaks “ideas” by “ideas” by the “ideas” by the “ideas” by the “ideas” by the believer.

However, the costs of the conclusion of meditation models may create quickly installed Cott Tokens. In a new paperKarnegie University Students University Students University officer LLM, recommends the training of LLM, and the developers offer overseeing the length of the breasts.

Length of control policy (LCPO) to ensure the correct response to the correct response to the correct response to the correct response to the correct response, as well as, “thinks of the” thoughts “in the preliminary budget. Experiments on LCPO-trained models can trade in a flat model between accuracy and costs, and they can repay large models on the abnormally amazing length. LCPO’s applications will help thousands of tokens to chat with LLM and to reduce costs to enterprises to interview the company’s applications.

LLM indicators lasts a length

Models like OpenAi o1 and Deepeek-R1 taught through comprehensive training (RL) to use Test-time interference Create cost trails before relying on the answer. As empirical evidence shows that the models are long, when the models lasted, they do better to think their thinking.

For example, at first, it was first prepared for pure rl without example from a person. One of the concept is the more effectiveness of this model, the more the model has been able to form a COT footage again.

In general, Long Clot Chain can lead to accurate answers, as well as computent in the use of scale models. Tens of thousands of thousands of tokens can be easily launched without making the test term monitor the testing budget and winning the test. There were several attempts to control the thinking chains, but they usually reduce the work of the model.

The length of the manufactured policy interpretation (LCPO) explained

Classical RL methodology to achieve the correct answer llms. LCPO has changed this paradigm: 1) Restrict a particular token length to get the correct results and 2) to get the correct results. Therefore, if the model has answered correctly, it will be punished, and it will be punished, and it will be punished, but small token, but a small token, but with the budget.

“Models taught by LCPO will learn to think of long-term restrictions, optimizing the performance performance and satisfy long-term restrictions,” wrote the performance performance.

The LCPPO suggested two damps: (1) LCPO inspection program and (2) more than (2) a targeted length of the LCPPO “

To test the technique, a model of 1.5b parameters of the researchers (QWEN-Disticle-Disticle-R1-1.5b) has been a better planning model (QWEN-Disticle-R1-1.5b) to create L1-MAx and L1-Max and L1-Max and L1 and accurate models. Based on mathematical problems with accurate and checked results. However, evaluation includes the problems of mathematics, as well as a mass understanding of the mass-scale language (Mmlu) Google-Proof Q & A (Gpqu).

Their results are interfering in short, effective, effective intervention, and long-term restrictions, with long-term restrictions on various length, with a short-term intervention and long-term restrictions on various length. Most importantly, on some assignments, L1 MODELs can increase the way the initial model model of the model in the original thinking to the following final budget.

Compared to S1 – another other method, which limits the length of L1 models, appears to 150% in various token budgets.

“This significant difference is made of two main factors.” (1) L1 often will be taught that the high quality of the highest quality of values ​​is taught, “he said.”

L1 is also overcome by 5% and GPT-4O to 2% to GPT-4O, and to 2% on the generation length. “Our knowledge is the best of the 1.5b model, and it can exceed the GPT-4O border models such as GPT-4O, such as GPT-4O,” like GPT-4O.

Interestingly, the model’s breast shows that he learns to regulate the reasonable thinking based on its basic budget. For example, in the future, the model can create related tokens related to self-correction and verification (ie “and drawing a picture” and “waiting”) and drawing.

In addition to standard reasoning of standard reasoning, the L1 model summarizes L1 Models Frialfer tasks, including Gpka and Mmlu.

In the models, these new research in their thinking will be able to use the modernization of enterprises by the real world and capable of large-scale climate. It is only a economically viable factor for complaints of larger, expensive models – the AI’s existence for complaints of high-scale and worldwide.

Have opened the discovery of researchers LCPO code and the Scales for L1 models.