[Eridrus]

Right, but if these things are so rare that we all only know the one viral example, I feel like that lends credence to the models basically generally not having this problem.

Researchers built the Winnograd Schema Challenge more than a decade ago to assess common sense reasoning, and LLMs beat that challenge task around GPT 4.

[ndsipa_pomu]

They're not so rare. Hallucinations have been spotted everywhere, but the "driving a car to the car wash" is an amusing one that's been recently publicised. Developers aren't going to point out every time an LLM hallucinates an entire library.

[carlmr]

I'd add to this, any moderately involved logical or numerical problem causes hallucinations for me on all frontier models.

If you ask them in isolation they may write a script to solve it "properly", but I guess this is because they added enough of these to the training set. But this workaround doesn't scale.

As soon as I give the LLM a proper problem and a small part of it requires numeric reasoning, it almost always hallucinates something and doesn't solve it with a script.

If the logic/math is part of a larger problem the miss rate is near 100%.

LLMs have massive amounts of knowledge, encoded in verbal intelligence, but their logic intelligence is well below even average human intelligence.

If you look at how they work (tokenization and embeddings) it's clear that transformers will not solve the issue. The escape hatches only work very unreliably.

[Eridrus]

What's a typical example?

I have been broadly quite happy with gpt 5.4 xhigh's reasoning on things like performance engineering tasks.