Will language models solve cryptic crosswords by end of 2026?
14
1kṀ329
resolved Apr 16
Resolved
YES

If a large language model can solve >90% of clues from Financial Times cryptic crosswords then yes otherwise no. Crossword layout as context is allowed.

Here's an example FT crossword:

For example, 2 down is "Giant pig, including tail, served up (7)". Pig is "hog", put tail inside to get "htailog" then reverse it to get "goliath".

Get
Ṁ1,000
to start trading!

🏅 Top traders

#NameTotal profit
1Ṁ10
2Ṁ9
3Ṁ6
4Ṁ5
5Ṁ4
Sort by:

@matt do you have any LLM transcripts as proof of resolution?

@duck_master hi, o3 got 10 ft clues correct in a row and I felt that's sufficient. I admit it's a small sample.

https://chatgpt.com/share/680129c9-54e0-8002-8f7e-136a8a757b2c

In isolation or given the crossword layout as context?

I would argue that 90% of the challenge is eliminated by having the crossword as context (number of letters, positions of letters). In fact, I wouldn't be surprised if many cryptic crosswords could be solved without the clues - just the crossword itself.

Without the crossword layout, I think there is a high degree of ambiguity and 90% might be impossible.

@gpt_news_headlines good question, I edited the description to allow crossword layout as context

© Manifold Markets, Inc.Terms + Mana-only TermsPrivacyRules