This year, many LRMs, who try to solve their first results, instead of spitting their first results. High scores in American Invitational Math Examination (Aime), a test given to 5 % of the top 5 % of American high school mathematics.
At the same time, a handful of new hybrid models that connect LLM with some facts checking system also created success. At the University of Sao Paulo, Brazil, the mathematician, Emily de Olivira Santos, points to Alfaproof of Google Deep Mind, which connects an LLM with a Deep Mind’s game playing model Alfazro, an important milestone. Last year, Alfaprov became the first computer program to compete in the silver medalist’s performance at the Olympiad Olympiad, one of the most famous mathematics competition.
And in May, the Google Deep Mind Model discovered better results to the model called Alfaovolo, which humans presented with no more than 50 soluble mathematical puzzles and many real -world computer science issues.
The increase in growth is clear. De Olivira Santos says “GPT -4 could not do more mathematics than undergraduate levels.” I remember that at the time of his release, it had to examine it with a problem in toopology, and he could not write more than some lines without being completely lost. “
Does this mean that such models are fully developed as a fellow Darpa hopes? Not necessarily, she says: “Mathematics Olympiad issues often include being able to carry out smart tricks, while research issues are highly exploited and often contains many, many dynamic pieces.” Success in solving one type of problem cannot reach the other.
Others agree. Mattered Martin Birdson, a mathematician at Oxford University, believes that the result of the mathematics Olympiad is a great success. They say, “On the other hand, I don’t mind this mind.” “This is not a change in the sense that ‘wow, I thought that machines would never be able to do so.’ I expected the machines will be able to do so.
The reason for this is that although the difficulties in the Olympiad Olympiad, as well as undergraduate tests like high school or Aime, are difficult, many of them have a model. “We have training camps to train high school children,” says Birdson. “And if you can train a large number of people to deal with these problems, why can’t you get training to do a machine?”
Sergey Gokov, a mathematician at the California Institute of Technology, who coaching mathematics Olympiad teams, has pointed out that questions do not change much between competitions in the style of questions. New problems are settled every year, but they can be solved with the same old tricks.