Less is more: UC Berkeley and Google unlock LLM potential through simple sampling

One New paper by researchers Google Research and University of California, Berkeley, A surprisingly simple test time demonstrates that a large language models (LLMS) increase their reasoning skills of the scale approach. Key? Sample-based search covers a technique to create numerous answers and use the model to use them.

The main finding can also download the main performance of the models outside the preview of the example of O1 in popular criteria, using random sampling and self-approval. Findings may have important impacts for business applications and protest the possibility that highly specialized training or complex architecture is always necessary to get the highest level performance.

Scale scales of current test time

The popular method for test time on the Test time on the LLMS is to prepare the model by strengthening the model to create longer answers with perceptions of thought (COT) prints. It is used in models such as approach Openai O1 and DeepSeek-r1. While useful, these methods usually require a basic investment in the training phase.

Another test time scale method “holds himself” here, where, where you choose a lot of answers and chooses the answer that is more often visible. While working in complex problems of self-consumption, the most recurring response is not correct.

Sample-based search offers a simpler and higher size for test-test scale: The model creates numerous answers and select the best through the inspection mechanism. Sample-based search Other test times are the unique advantage of whether they are written in their papers and writing on researchers paper.

More importantly, the sample-based search can be applied to any LLM, including not trained in the mindset.

Sample Based Search Works

Researchers use a language model to create candidate answers, researchers are aimed at a minimalist application of sample searches. This is the “self-confirmation” process that the model is assessed by the external basic truth or relying on symbolic validation systems.

The algorithm works in a few simple steps:

1-Algorithm, using a language model, begins with a number of candidates with a candidate solution. This is done using a non-zero temperature to give the model several times and collect various answers.

2-The answer of each candidate is subject to the verification process in which the LLC is asked several times to determine if the answer is correct. Average to create the final verification score for the answer after the verification results.

3- The algorithm chooses the highest goal in the last response. When many candidates are at close range of each other, LLM is asked to compare them in pairs and choose the best. The response to the comparison with the pair is selected as a final response.

Researchers reviewed two main baltages for test time:

Example: The number of answers generated for each entry problem of the model.

Verification: The number of calculated verification points for each generated solution

How sample-based search compares with other methods

The study revealed that it was improved by terminal-based search, even the test time calculation account was scaled more than the point of a very grown point.

On a sufficient scale, this minimalist implementation, aim and mathematics significantly increases the views of criteria. For example, the performance of Gemini 1.5, O1-Preview and Gemini 1.5 Flashes, which are clearly prepared on the problems of thinking problems, exceeded the Gemini 1.5 Pro.

“This emphasizes the importance of sample search for not only measurements, as well as a simple basis for comparing other testing strategies and measuring models in search capabilities,” researchers write.

Although the outstanding qualifying results are impressive, the costs may also be prohibited. For example, with 200 samples and 50 verification steps, a survey from AIME will produce a $ 1,300-dollar to $ 650 with a survey, Gemini 1.5 Pro. However, this is a very minimal commissioning approach to the sample search and is compatible with the optimization methods offered in other studies. Conference costs with Smarter sample and inspection methods can be significantly reduced By using small models and Create fewer verses. For example, using Gemini 1.5 Flash to verify, the expenses go down to $ 12 per question.

Effective self-approval strategies

There is an ongoing controversy that LLMS cannot verify their responses. Researchers have identified two main strategies to improve self-approval using the calculation of the test time:

To compare the answer candidates directly: Disagreements between candidate solutions are strongly showing potential mistakes. By providing multiple answers to compare, the model can better identify errors and hallucination by solving the main weakness of the LLMs. Researchers describe it as an example of “hidden scale”.

Task Speasing Rewrite: Researchers suggest that an LLM’s optimal way of exit depends. The chain of thinking is effective to solve thinking tasks, but the answers are easier to check when it is written in a more formal, mathematical style. Verifying can rewrite a more structured format (eg, theorem-Lemma-proof) before appreciating the candidate answers.

“Models are waiting for the model self-confirmation for rapidly improving, because models learn to use the principles of secret scale and output.”

Effects for Real World Applications

The study shows that a relatively simple technique can achieve effective results, complex and expensive model architecture or the need for training regimes.

It is also a technique that is expanding, allows enterprises to increase the performance to select more calculation resources to select more calculation sources. In addition, the developers allow the front language models to push outside the restrictions on complex tasks.

“This allows other test time to compile calculation strategies, parallable and arbitrary scale, because it is increasingly large with comprehensive calculation budgets, we expect to be resolved with increasingly large computing budgets,” researchers write.