It isn't the first time that an AI has claimed sucess in a coding competition, but this one is impressive and highlights what the coming AI revolution my be all about.

The International Collegiate Programming Contest (ICPC) is an annual competitive programming competition among the universities of the world that culminates in a grand final. This year (2025) it was held in Baku Azerbaijan and had an interesting additional competitor in the form of Google's Gemini 2.5 Deep Think. You can use a "lightweight" version of the same program if you are a Google AI Ultra subscriber. There isn't much information on what modifications were made to the competitor version, but I think we are fair in calling it "heavyweight" - whatever that means.

So confident was the entry that it gave the puny humans a whole 10 minutes head start. Not such a huge margin in a  three-hour contest but it did manage to solve 8 of the 12 problems in 45 minutes. In the full three hours it managed 10 of the 12. This entitled it to a gold medal, which only four of the 139 human teams won. It is also worth pointing out that the team from the St Petersburg State University solved 11 of the problems. 

The most interesting of the 10 problems that the AI solved was problem C. This was solved by none of the human teams. The problem is one that needs knowledge of linear programming to solve a flow network type problem:

"The story continues! For several years now, your town has been gifted with an abundance of Flubber, the adorable-but-slightly-flammable-and-toxic-and-acidic-and-sentient-and-mischievous man-made chemical. The search continues for more (or, well, any) uses for the substance. But in the meantime, the Flubber factory continues to produce it at full capacity. Efforts to shut it down have failed, partly because nobody is sure who is actually running the factory."

Well that might have confused a non-film watching AI ... and to be honest the rest of the explaination of the problem would take some studying to understand fully. Overall, I can say it left me flubbergasted.

The basic idea is that you have a number s of producers or stations, r is the number of reservoirs and d is the number of ducts connecting the other two. The output of the factory flows into station 1 and then flows down the ducts into other stations and reservoirs. The first line of the problem consists of s,r,d and there follows one line for each duct starting with i the station that the duct drains and n the number of outputs of the duct. There then follows n pairs o and p that give the station or reservoir that the duct drains and the percentage flow.  For example:

s r d
2 3 3  two stations, three reservoirs and three ducts

i n o p  o p
1 2 3 80 4 10   
duct drains 1  with 2 outputs one to 3 at 80% and one to 4 at 10%
and so on... 

1 2 2 40 4 30 
2 1 5 100

which defines the network:

OK, so some of you may have got that on first reading, but not me. The problem is to find the percentage f that is the highest that, for some configuration of drainage all reservoirs, receives at least f% of the total. Notice that the only degrees of freedom you have in the solution is the outflows from the stations. So, for station 1 you could put 10% to the left and 90% to the right. This would give you 8% into 3 and 9% into 4 and if station 2 passed 100% on then 5 would get 40%. Thus the f in this case is 8% and you have to adjust the flows to make this as large as possible.

Good luck programming that. 

It turns out that Gemini solved it thus:

"Gemini found an effective solution with a clever insight: it first assumed each reservoir has a "priority value" representing how much each reservoir should be favored compared to the others. When given a set of priority values, the best configuration of the ducts can be found using a dynamic programming algorithm. Gemini discovered that by applying the minimax theorem, the original problem can be approached by finding the priority values that make the resulting flow most constrained. Leveraging the relationship between priority values and optimal flows, Gemini used nested ternary searches to quickly find optimal priority values in the bowl-like convex solution space, and solved Problem C."

OK - well I really don't know what a "priority value" is in this case, but think I could get closer to the idea with some experimenting. You need to keep in mind that none of the humans solved the problem, but Gemini solved it in 30 mins which is about average for the problems it solved - so no sweat there then. 

Now compare this to a problem that the humans found relatively easy, but Gemini failed to solve:

"To help her elementary school students understand the concept of prime factorization, Aisha has invented a game for them to play on the blackboard. The rules of the game are as follows. The game is played by two players who alternate their moves. Initially, the integers from 1 to n are written on the blackboard. To start, the first player may choose any even number and circle it. On every subsequent move, the current player must choose a number that is either the circled number multiplied by some prime, or the circled number divided by some prime. That player then erases the circled number and circles the newly chosen number. When a player is unable to make a move, that player loses the game. To help Aisha’s students, write a program that, given the integer n, decides whether it is better to move first or second, and if it is better to move first, figures out a winning first move."

Could it be that Gemini isn't clear what a prime is - or is it just arithmetic? Surely not. The only other problem Gemini couldn't solve took the humans ages - well 270 minutes, which is well over the average. The problem involved spatial reasoning and even the humans aren't as good at that as logic.

Does this mean that Gemini thinks differently?

Without more information it is impossible to say, but the humans were impressive as well as the AI.

There is much more information on the project's GitHub page. Perhaps more analysis would be useful.

More Information

Gemini achieves gold-medal level at the International Collegiate Programming Contest World Final

49th ICPC World Finals Baku

https://github.com/google-deepmind/gemini_icpc2025

Related Articles

Human Programmer Outwits OpenAI's o3

The Single Issue Of 2025 - AI

AI Wins At Rubik's Cube With Just One Hand

AI-Based Artwork Wins Prize

To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on Twitter, Facebook or Linkedin.

Comments

Make a Comment or View Existing Comments Using Disqus

or email your comment to: comments@i-programmer.info