I was discussing this with a grandmaster, here are some ideas we came up with:
- Remove solve counts. Solve counts basically function as a communication method for cheaters during the contest. Once some cheaters realize a problem is solvable by AI, more and more will follow.
- Force people to solve problems in order during contests. To unlock problem B, you must solve problem A. To unlock problem C, you must solve problem B. You get the idea. This makes it so that once cheaters hit a problem that cannot be solved by AI, they're doomed for the rest of the competition. This would make it so that the problem committee can set traps. They can engineer one or two problems per contest that break AI (either logically, visually through the use of complex diagrams, or textually by embedding hidden messages designed to derail LLMs) then investigate suspicious or new accounts that failed to get through those problems. This brings me to my third and final point.
- More creative problem design; less contests, higher quality problems. Maybe the fact that AI is getting so good, so fast, at competitive programming problems means most problems are too generic. I've seen problems in the past that AI can't solve, and I even came up with my own yesterday. I know it's possible.
What do you think MikeMirzayanov? As I've stated before, I and (most likely) many others would be willing to help with development for new features. The pressure is not all on you to fix things, we are a community.
