Blog
Notes on meaning, knowledge, and game characters.
-
The curse of knowledge in role-playing language models
Language models that speak as game characters repeat what those characters could not have heard or understood, even when they can say exactly what the character perceives.
-
Some failed ideas in semantic parsing
Four ideas for DRS and AMR parsing that did not survive a matched baseline — RL post-training, a logic-verified reward, a translation-invariance reward, and DRS as a meaning checker for machine translation.
-
From a model that talks to a character that lives
Introducing NPCBank, source-grounded game characters with inspectable dialogue cases, and three challenges for building characters that adapt to their world without losing themselves.
-
OntoURL: A Benchmark for Ontological Understanding, Reasoning and Learning
A benchmark of 58,981 questions over 40 ontologies, testing whether language models can do more with symbolic knowledge than recognise it.
