Apple Machine Learning Research has a new preprint: “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity.” [Apple, PDF] “Lar…
Further support for the memorization claim: I posted examples of novel river crossing puzzles where LLMs completely fail (on this forum).
Note that Apple’s actors / agents river crossing is a well known “jealous husbands” variant, which you can ask a chatbot to explain to you. It gladly explains, even as it can’t follow its own explanation (since of course it isn’t its own explanation but a plagiarized one, even if changes words).
I think what I need to do is to write up a bunch of puzzles, assign them randomly to 2 sets, and test & post one set, while holding back on the second set (not even testing it on any online chatbots). Then in a year or two see how much the set that’s public improves, vs the one that’s held back.
That would be the best way to actively catch the cheating happening here, given that the training datasets remain confidential. But I also don’t know that it would be conclusive or convincing unless you could be certain that the problems in the private set were similar to the public set.
In any case either you’re doubledipping for credit in multiple places or you absolutely should get more credit for the scoop here.
I’d just write the list then assign randomly. Or perhaps pseudorandomly like sort by hash and then split in two.
One problem is that it is hard to come up with 20 or more completely unrelated puzzles.
Although I don’t think we need a large number for statistical significance here, if it’s like 8/10 solved in the cheating set and 2/10 in the hold back set.
Further support for the memorization claim: I posted examples of novel river crossing puzzles where LLMs completely fail (on this forum).
Note that Apple’s actors / agents river crossing is a well known “jealous husbands” variant, which you can ask a chatbot to explain to you. It gladly explains, even as it can’t follow its own explanation (since of course it isn’t its own explanation but a plagiarized one, even if changes words).
edit: https://awful.systems/post/4027490 and earlier https://awful.systems/post/1769506
I think what I need to do is to write up a bunch of puzzles, assign them randomly to 2 sets, and test & post one set, while holding back on the second set (not even testing it on any online chatbots). Then in a year or two see how much the set that’s public improves, vs the one that’s held back.
That would be the best way to actively catch the cheating happening here, given that the training datasets remain confidential. But I also don’t know that it would be conclusive or convincing unless you could be certain that the problems in the private set were similar to the public set.
In any case either you’re doubledipping for credit in multiple places or you absolutely should get more credit for the scoop here.
I’d just write the list then assign randomly. Or perhaps pseudorandomly like sort by hash and then split in two.
One problem is that it is hard to come up with 20 or more completely unrelated puzzles.
Although I don’t think we need a large number for statistical significance here, if it’s like 8/10 solved in the cheating set and 2/10 in the hold back set.