xuan (ɕɥɛn / sh-yen)

xuanalogue.bsky.social

22 February 2025 08:49

My own bet is that we can scale model-based Bayesian approaches to ToM reasoning to much more open-ended settings than they've been historically applied to (a topic I've given talks on recently, if you want to hear more about it)!

xuan (ɕɥɛn / sh-yen)

xuanalogue.bsky.social

22 February 2025 08:49

In other words, there's still lots of work to be done if we want AI assistants that can reliably infer & reason about the goals & beliefs of humans they're interacting with -- current reasoning training paradigms don't give it for free, and scaffolding helps but isn't a cureall!

xuan (ɕɥɛn / sh-yen)

xuanalogue.bsky.social

22 February 2025 08:49

It was fun helping out on this paper! Haven't worked on nat lang ToM benchmarks in the past, and it was cool to see that SMC-like scaffolding of LLMs works better than o1-style reasoning models -- though both are still less reliable than explicit model-based ToM reasoning.

Tomer Ullman

tomerullman.bsky.social

14 December 2024 08:14

thinking of calling this "The Illusion Illusion" (more examples below)

Leshem (Legend) Choshen @EMNLP

lchoshen.bsky.social

14 December 2024 08:14

Models can self-improve🥷 by knowing they were wrong🧘‍♀️ but when can they do it? Across LLM families, tasks and mechanisms This ability scales with pretraining, prefers CoT, non QA tasks and more in 🧵 alphaxiv.org/abs/2412.02674 @yus167.bsky.social @shamkakade.bsky.social 📈🤖 #NLP #ML