Блог пользователя AlterAB

Автор AlterAB, история, 17 месяцев назад, По-английски

Seeing everyone achieving titles they want with AI, is making CP less fun during contests. Feeling happy that I implemented a particular logic during contest and then seeing that I still was not able to cross the rating barrier I want to break. After each contest I check how other people near me implemented that, turns out the 30-40% of them all are implementing the problem in one way fashion resembling GPT with minor modifications in formatting and change in data structure wherever they can. I want the criteria of AI plagiarism detecting to become way stricter than it is now, like if caught cheating you are banned, why is there leniency towards cheating by giving them warnings, they should not be cheating here from the first place rather they should be focusing on how they can improve and upskill.

  • Проголосовать: нравится
  • +120
  • Проголосовать: не нравится

»
17 месяцев назад, скрыть # |
 
Проголосовать: нравится +34 Проголосовать: не нравится

It's crazy, everyone is suddenly a genius.

»
17 месяцев назад, скрыть # |
 
Проголосовать: нравится +32 Проголосовать: не нравится

I wouldn't even get top 100 in today's div 3 (wasn't able to do H during contest and took 50 min to come up with a solution for G). It's just sad at this point ;-;

»
17 месяцев назад, скрыть # |
 
Проголосовать: нравится +3 Проголосовать: не нравится

There is a certain dilemma with the matter of banning based on ai plagiarism. It's like, what criteria could you even use to say that a single submission is definitively ai made? I mean, of course there are many, but also a lot of them can be avoided by just rewriting the code in one way or another.

And then there is the matter of false positives. One of the people I was coaching got skipped on some submission/was reported. I checked the skipped submission and it seemed fine to me: same style they had been using, idea wasn't that out of the way, code was short, etc, something that I believed they could reasonably solve/program in the given time frame. When I compared it to other submissions, it became clear to me that although the code was different, internally they all had the same structure. And while that is suspicious, the problem code was also relatively short and the mathematical model very straightforward.

Because of that, I think that false positives could be something that can be managed by forcing more complicated solutions where the models we all develop + consequent implementations would necessarily have to be different, but there is a limit even to that.

So, the most reasonable way to resolve this problem seems to be to implement an "advanced" trust system as discussed by many previous blogs here on cf. Managing trust could also help if/when llms start exhibiting even less distinguishable characteristics.

  • »
    »
    17 месяцев назад, скрыть # ^ |
     
    Проголосовать: нравится 0 Проголосовать: не нравится

    frankly I think top 200 in this div 3 is full of cheaters like atleast >50%... like how is the following performance possible many have a consistent 3000-4000 in div 2/div 3 contests and suddenly in this contest are in top 200... many people I have seen could not even seem to solve question a and suddenly started spam solving all questions

»
17 месяцев назад, скрыть # |
 
Проголосовать: нравится +8 Проголосовать: не нравится

I have a very simple solution, which i came up by seeing yesterdays C

Many people have chatGPT code and there are few tests which give tle on those. Many of them have been hacked during the hacking phase.

It worked cause the constraints were very clever. It said n * m <= 1e5, didnt say the limits of n and m. I believe this fooled gpt into thinking maybe a n * n * m or n * m * m sol works which is why so many people have hacked solutions for C

So we dont need to change the problems entirely, let them be solved by AI, after the contest it will be hacked or gone by system testing. Eventually people will get frustuated and stop using it

  • »
    »
    17 месяцев назад, скрыть # ^ |
    ← Rev. 2  
    Проголосовать: нравится 0 Проголосовать: не нравится

    funny thing is even n*n*m will work if you just take the transpose and take the minimum n... So might not work everytime... similar idea to an atcoder abc 410 f

  • »
    »
    17 месяцев назад, скрыть # ^ |
     
    Проголосовать: нравится +3 Проголосовать: не нравится

    Yes this is so sad, people are GPTing so blindly and not getting caught. The failure on system testing on C clearly showed the people who GPT, it would be great to see if authors can take this into consideration for future contests

»
17 месяцев назад, скрыть # |
 
Проголосовать: нравится +3 Проголосовать: не нравится

My classmates, they have no training, do you believe that they solve 7 problems in the div3 yesterday. What's more, they copy each other and just modify the name. It's a shame!