logoalt Hacker News

YeGoblynQueennelast Monday at 10:54 PM2 repliesview on HN

LLMs don't search trees. They generate plausible proofs and a human has to check it's true. Repeat until.


Replies

tptacekyesterday at 12:51 AM

That's not what happened here. This isn't a proof; it's a counterexample. The model was perfectly capable of verifying its correctness. You could have verified it by hand if you wanted; the verification is trivial. Finding it was the hard part.

show 1 reply
astrangelast Monday at 11:13 PM

Agents can generate formal proofs that are checked with an oracle like Lean and can run in a loop.

show 1 reply