We ground this distinction in cognitive neuroscience.

Years of empirical work show that humans have specialized neural machinery for language processing (reading, listening, speaking, etc), which is distinct from brain mechanisms underlying other cognitive capacities (social reasoning, intuitive physics, logic and math…) 3/

Bonus: a preliminary exploration of #ChatGPT responses shows that it might also have an impossible-implausible gap (although a more detailed investigation is of course needed).

9/end

- explicit plausibility information emerges in mid LLM layers and then stays high
- implausibility signatures generalize poorly across animate-inanimate (impossible) events and animate-animate (unlikely) events
- a probe trained on both active and passive voice sentences is as successful as a within-voice probe (but a probe trained on only one voice type fails to generalize)

6/

- LLMs generalize very well between active and passive versions of the same sentence BUT not as well as humans for synonymous sentences (The teacher bought the laptop / The instructor purchased the computer).

5/

In follow-up tests, we show that
- LLM scores depend both on plausibility and surface-level factors like word frequency (meaning that distributions for plausible and implausible sentences are highly overlapping)

4/

LLMs are almost perfect when assigning likelihood to possible vs. impossible events but aren’t as good when it comes to likely vs. unlikely events.

(our baseline language models also show this effect)

3/