Very little: UC Berkeley and Google Inlock LLM the ability through simple sapping

A New sheet and researchers from Google research with California University, Berkeley, It shows that the process of reduction of testing test can promote the skills of major languages ​​(LLMS). Key? To set up the search-based search, which process depends on multiple responses and use the form to prove them.

Access to the applicant is that although the stabilizing sessions, using Sappping random, can lift the function of the Business and criticism of a special or conventional education is always important to achieve a great job.

The limit of a long study study

The most common way to find the test examination of the Llms and training the color through the promotion to learn to associate with a long mind and mind (cot). This process is used for colors such as Tsetai O1 with Duuceseek-r1. Despite the profit, these methods usually requires a lot of money in teaching.

Another way to submit a test and “self-choice”, “when the model gives several answers to this question and select the solution when it is used difficult problems, such as this, the repeated answer is not correct.

The extent of the extent of the extent of the test provides a simple and lowest way to reduce multiple responses and select the best option through verification machines. Examples search may change other experimental methods and as researchers wrote in their papers, “It also has the opportunity to be equal to shame: just letting the answer: Just answer many questions.”

The most important search, examples will be used to each llm, including those who are not clearly trained.

How to search for samples

The researchers look very much in a temporary plan, using a language to all the candidates and verification. This is a form of identifying, where the model examination of what they do without trusting with the points of the facts to talk to or the sophistication.

Algorithm works a little bit:

1-algorithm begins to make a solution to the problem in the problem using words. This happens to give the same type of range and use of the temperature of zero to make different answers.

An individual response makes the confirmation of the LLM confirmed several times to know if the answer is correct. The result is designed to make the final reflection.

3- algorithm chooses the highest response as the last answer. If candidates to be unhappy with each other, LLM is encouraged to be pretended to be with the best choice. A solution that wins a lot of comparison to be selected for the final answer.

The researchers saw two attempts to test exams:

Sampling: The number of answers that best describes each problem.

Verification: Number of authenticated figures that are made for each answer

How the sample search is similar to alternatives

The surveys showed that performances continue to control the sample search, even when experimental periods can be decreased that will make things out of the best.

At enough extent enough, install this minimical composition composer accurately to accurate discussions as representatives and math. For example, Gemini 1.5 Pro

“This does not indicate the importance of analysis, as well as verifications for the required amount of time you can match the switch,” research scores.

It is important to note that even the results of the research are interesting, the price can be restricted. For example, with the examples of 200 and a definite 50-session of each seat, a question from Aime produces a million tokens, which cost $ 650 and gemini 1.5 pro. However, this is the smallest-looking way, and it is in agreement with advertisements that are asked in other subjects. It’s a solution and verification methods, the amount of income can be very low by Using small varieties with To make a few tokens. For example, using quni 1.5 to verify, the cost drops up to $ 12 for each question.

The way to do

There are some dialogs that’s there if the llm can verify their answers. The researchers have identified two of the directions to show you using the test method:

To evaluate directly to answer: The conflicts between the combined answers to indicate strongly that can happen. By providing the show with a number of answers to match errors and compares, the weakness of the llms. The researchers tell this as an example of “weakness.”

Special Writing: The researcher thinks that the correct LLM’s nationality is based on the work. The ideas – a-concepts are helpful to resolve the discussion, but the answers it’s easy to confirm when you write a common type. The testimonies can rewrite the ideal responses to be successfully made (eg, the AMROM-tabernacle).

“We look forward to a better self-accuracy of the best of time, as examples learn to reduce the Strectic Principles and releases, and driving well in search,” wrote.

The result of real work

The lesson indicates that a simple training process can achieve interesting, which may also reduce the importance of a complex and scarce.

This is a way to get the burning, contributor of adding to the functioning of work using herbs and certifications. It also helps makers to choose the lines of more than the bounds in difficult tasks.

“Since it meets other testings, matching and makes it possible, and we look forward to the most important search, we have a great budget,” research budgets, “investigated.