Пожалуйста, подпишитесь на официальный канал Codeforces в Telegram по ссылке https://t.me/codeforces_official. ×

Solving 1137D correlates negatively with your old rating

Правка en4, от ma_moutahid, 2026-11-22 17:27:25

Solving 1137D correlates negatively with your old rating

Instead of making observations where I needed to today (problems DEF), I noticed that some IOIers and high-peak-rating acquaintances of mine solved E but struggled with D, even though the submission counts point to D being the easier problem. So I decided to dig deeper.

The setup

Round 1137 (Div. 2) had 2878 official participants. D was solved by 721 people and E by 231. D and E were both worth 2000 points. Two groups solved A, B, C and E:

  • ABCE: no D, 45 people. Strictly fewer problems.
  • ABCDE: 166 people. Strictly more problems.

Then I asked the Codeforces API one question: what were these people rated before?

Rating on Nov 22 of… ABCE (fewer problems) ABCDE (more problems) Gap p (permutation)
2022 1736 1393 +343 0.04
2023 1768 1298 +469 <0.001
2024 1772 1523 +249 0.02
2025 1763 1646 +116 0.04
Today (before the round) 1726 1865 −139 –

Medians among accounts that already had a rating at that date.

Every year from 2022 to 2025, the people who solved fewer problems today were rated higher. Then, as of this year, the order flipped.

Rating history of ABCE vs ABCDE solvers

The ABCE group has been flat for four years. The ABCDE group climbed steadily, starting around January 2025, which, as I'm sure everyone knows, is when everyone collectively got smarter for no particular reason.

Other things that are surely a coincidence

  • 42% of the ABCDE group created their accounts after Nov 2024, vs 20% of the ABCE group.
  • Averaged over each account's whole lifetime, the two groups are level (1703 vs 1685). The difference isn't who is stronger, it's when they got strong.

"But maybe this happens every round"

I ran the same split on other Div. 2 rounds: the first Div. 2 of Jan 2025 and of Jun, Jul and Aug 2026, plus the two most recent rounds where D and E were comparable. (The Sep and Oct rounds had split problems like C1/C2, so the split doesn't apply cleanly.) Same metric: median rating 2 years before the round, ABCE minus ABCDE.

Round D/E points D/E problem rating D/E solves ABCE ABCDE Gap (2 years before)
996 (Jan 2025) 2250/3000 2000/2800 780/2 0 1 –
1102 (Jun 2026) 1750/2500 1500/2100 4070/354 14 321 –
1108 (Jul 2026) 1750/2250 1900/2600 959/27 5 21 –
1113 (Aug 2026) 1750/2500 1400/2000 3115/633 3 611 –
1045 (Aug 2025) 2000/2250 2300/2300 387/233 179 49 −17 (p=0.55)
1118 (Aug 2026) 2000/2250 2500/2400 53/99 58 11 −353
1137 (today) 2000/2000 TBD 721/231 45 166 +249 (p=0.02)
  • Normal rounds: D is much easier than E, so "solved E but not D" barely exists.
  • D and E equally hard (1045): the two groups have the same history.
  • D genuinely harder for humans (1118, fewer solves and a higher rating): solving D marks you as stronger, as it should.
  • 1137: the only round where solving D marks you as less established, while D was solved by three times as many people as E.

Caveats, before the comments write them for me

  • 45 vs 166 is not a big sample. Also, n=1 suspicious round.
  • People do improve. Some of those +400 climbs are real, and I'm not accusing anyone in particular.
  • Being able to solve E but not D could just be a skill profile that happens to come with older accounts. Then again, the 1118 row argues against that.
  • If you solved ABCDE and were 1300 in 2023: congratulations, sincerely. You're the reason this post has a caveats section.

Still, a problem being much harder for one kind of contestant than for another tends to move ratings in exactly this direction. Namely, and I suspect I fall into some of these categories myself, plateaus, overfitting and losing shape.

This is not an "AI is ruining Codeforces" rant

I do competitive problem solving for fun, and hardly any circumstance can flip how I feel about it be it cheaters or else. Though I'll admit it's more fun when I know I'm up against other honest people who love the game as much as I do. As for careers, I already consider myself in a race against AI, so I don't mind having a free benchmark for my progress on that front.

In any case, this round woke me up to a few things that have nothing to do with whether anyone is cheating:

  • My skip-meter needs work. E was right there, worth the same points, and I never got to it. I need to get better at deciding when to drop a problem and look at the next one.
  • My training is overfitted. I need more variety in where my practice comes from, especially virtual contests, so I don't get too used to any one format or problem style.

If you have suggestions on either front, or think I misread something about this round, I'd love to hear them in the comments.

All data from the Codeforces API. Ratings 2 years back are reconstructed from rating changes.

История

 
 
 
 
Правки
 
 
  Rev. Язык Кто Когда Δ Комментарий
en5 Английский ma_moutahid 2026-11-22 17:30:14 115
en4 Английский ma_moutahid 2026-11-22 17:27:25 0 (published)
en3 Английский ma_moutahid 2026-11-22 17:26:54 227
en2 Английский ma_moutahid 2026-11-22 17:21:58 180
en1 Английский ma_moutahid 2026-11-22 17:17:33 5086 Initial revision (saved to drafts)