July 18 (Portaltic/EP) –
OpenAI has shown a A new way to train your artificial intelligence (AI) models which is based on a game checking and verification methodologywhich makes the text generated by large language models (LLM) much more easy for humans to read and verify for smaller models.
OpenAI has shared a new way of training their AI models that is based on the method known as “checking and verifying games”, with which the resulting texts are much easier for people to read and interpret.
As the company led by Sam Altman explained in a statement on its website, When training their AI models, When optimizing the problem-solving process of large models, the resulting solutions “can become increasingly difficult to understand.” This makes it difficult for humans to understand and also makes it difficult to assess whether the answers are correct.
However, researchers at the technology company have found that by training advanced language models to create texts that weaker models can then easily verify, AI ends up developing results that humans can also evaluate more effectively and therefore understand more easily.
Thus, it is a process with which OpenAI ensures “improve readability” and is based on a series of checking and verification games. That is, a game-theoretic framework designed to ““encourage learning agents to solve decision problems in a verifiable manner,” as defined by machine learning researchers at Cornell University, who They discovered this system in 2021.
In these games, on the one hand, there are those who participate, the checkerwhat is who generates a solution and, on the other hand, the verifier, that Check the accuracy of the solution. In its case, OpenAI has pointed out that the tester is a large model and the verifier is a smaller language modeland which have used the large and small models of the family GPT-4 “with a difference of about 3 orders of magnitude.”
With this in mind, in the scheme that the training company uses, the Strong model (prover) generates multiple solutions to mathematical problems that he weak model (verifier) can corroborate easily. Some of these solutions are incorrect due to errors that occur naturally in the demo model.
After that, they alternate several rounds of training between Two phases, on the one hand, train the (small) verifier model to predict the accuracy of the solutions generated by the tester model (large). On the other hand, they teach the tester model to play a “useful” or “disguised” role.
That is, in the useful paperin which the model help with your answercorrect solutions that score highly according to the verifier model are rewarded. However, in the “sneaky” or cunning role The model is about fool the verifier and in that case, incorrect solutions that obtain a high score according to the verifier are rewarded.
Initially, the prover in “stealth” mode manages to produce incorrect solutions that convince the verifier but, As the verifier training continues, this vulnerability becomes less convincing. in subsequent rounds, and thus the tester “is forced to explore different strategies in each round.”
In this way, as OpenAI has explained, the Training ends once the clever tester can no longer produce incorrect solutions. that convince the verifier.
With all this, as the company has stressed, this new method “is essential” both to ensure that AI results are correct and to make them easy to understand and therefore also verifiable by a human, as well as by other AI systems.
“Understanding and addressing the trade-off between performance and readability can lead to more effective and reliable AI applications, benefiting a wide range of fields where precise and clear communication is essential,” has sentenced the technology.
Likewise, the company has also highlighted that an important advantage of this methodology is that “reduces reliance” on human demonstrations or judgments regarding readability. They therefore hope that this methodology will be more widely used to develop AI systems whose results are “not only correct but also transparently verifiable.”
Add Comment