[fauigerzigerk]

If a driver is tasked with visiting a number of places, they will probably choose a reasonably good route. If the driver claims to have found the optimal route, it may not be true, but it's still not a hallucination and it's still a pretty good route.

The driver certainly cannot be relied on to always find an exact solution to an NP-complete problem. But failure modes matter. For practical purposes, the driver's solution is not simply "false". It's just suboptimal.

If we could get LLMs to fail in a similarly benign way, that would make them far more robust without disproving what the posted paper claims.