Deepmind claims that the EU is better than gold medalists of the International Math Olympiad

Google Deepmind, the AI ​​system developed by Google’s leading AI research laboratory, the average gold medalist in the international math competition in solving geometry problems.

A system called AlphaPeometry2, an improved version of a system, Alifhometry, This Deepmind was released in January last. One New edition workDeep researchers behind Alphajometry2 can solve 84% of the EU in the last 25 years, 84% of geometry problems, a math competition for high school students in a math competition.

Why does Deepmind consider a high school-class math race? Well, the laboratory thinks that the key is lying in the detection of new roads to solve the key to hard geometry problems – specifically Geometry Problems in Evkli.

Prove mathematical theorems or requires the ability to choose one of a number of justifications and a number of possible steps to solve a teerem (eg the pythagoru theorem). This problem can make solvents – if Deepmindin fee – to be a useful part of future general purpose AI models.

Indeed, a system that connects the passenger with Alphapoofoof last summer, Deepmind, Alphajomemetry2, a system combined with the AI ​​model to solve four problems from 2024 IMO to the formal mathematics. In addition to geometry problems, such approaches can be extended to other areas of mathematics and science – for example, with complex engineering calculations.

Alphajemetry22, EU models have several main elements, including Google’s Gemini family and a language model called “Symbolic Engine”. The Gemini model helps a symbolic engine using the mathematical rules to infect solutions to problems, to come to a certain geometry theorem.

The Olympics are based on diagrams needed “Construction” to be added before solving geometry problems, points, lines or circles. The Gemini model of Alphajomemetry2 predicts which units can be helpful to add a diagram referred to for the engine allocations.

Basically, Alphajemetry2’s Gemini model offers a form of official mathematical steps and structures in the engine that complies with certain rules – these steps check these steps to follow these special rules. Search algorithm, Alphajeometry2 allows you to carry multiple searches for solutions in parallel and conduct many searches to maintain useful findings in a common knowledge database.

Alphajemetry22 considers the proposals of the Gemini model to a “solved” problem when it comes to a proof of symbolic engine known principles.

Thanks to the complexity of translating support in the format, AI can understand, there is a continuation of a usable geometry information. Thus, Deepmind, more than 300 million theorems and various complex evidence created their synthetic information to produce the alfajometry2 language model.

The Deepmind team has chose 45 geometry problems from IMO competitions (since 2024) in the last 25 years. Then, “translated them” has become a greater bunch of 50 problems. (Some problems for technical reasons should be divided into two.)

According to the paper, Alphajometry2, 40.9’s medium gold medalist solved 42 of 50 problems that cleaned the score.

There are restrictions. Something techniques prevent alifhometry2 from solving problems with a variable number of dots, nonlinear equations and inequalities. And not alifhometry2 technically The first AI system will reach a gold-level performance at the geometry level to achieve this with a solution to this size.

Alphajomemetry2 was worse in another set of hard imo problems. For an additional problem, the Deepmind team chose the problems – a total of 29 candidates for IMO exams by mathematics, but this has not yet visible in a competition. Alphajometry22 can solve only 20 of them.

Again, the results of work are likely to be controversed in the symbol manipulation of AI systems – that is, using the rules, manipulates symbols with neural networks such as brain.

Alphajetry2, a hybrid approach accepts: His twin model has a nervous network architecture, if the symbolic engine is in order.

Supporters of Neuron Network Techniques can occur in the imagery generation, the recognition of speech, massive amounts of information and anything else. As a line editing of symbolic systems, the word processor Software, as a line editing of the Scriptural Software, which solves the sets of symbol-manipulation rules dedicated to special work, is trying to resolve the tasks through statistical rapprochement and examination.

The foundation stone of strong AI systems such as neural networks Openai’s O1 “justification” model. However, the symbolic EU supporters claim, not all; Symbolic AI, coding the world’s knowledge effectively, encourage these ways to effectively codify ways through complex scenarios and “explain” how they arrive.

“This type of criteria, such trends and this issue, including language models, including” justification “, including” justification, “is striking to continue the fight against a number of simple teams.” Vince Conitzer briefed at the University of the University of Computer Science Techcrunch. “I don’t think it’s all smoke and mirrors, but I do not know what the next system is waiting for. This system is likely to be very effective, so we need to understand the risks they are urgently and better pose.”

Alphajomemetry22 Perhaps the two approaches – characters show that the manipulation and nerve networks – combined It is a way forward to the generalized AI search. Indeed, according to Deepmind paper, O1, which is a neural network architecture, could not solve any of the IMO problems that AlphapingMeMetry2 could answer.

This may not be the case forever. In paper, the Deepmind team, Alphajometryry2’s language model is primary evidence that the language model will be partial solutions without symbolic engine, he said.

“((The) SUPPORTS Support opinions” Wrote as symbolic engines), “The Deepmind team wrote as the team”, but the speed of (model) has been improved and hallucinations Absolutely resolved, tools will remain important for math applications. “