LLMs may pick up on the difference between ‘we’re friends’ and ‘we’re just friends’
The author says human-LLM comparisons, generation experiments and mechanistic interpretability methods show how one word shapes a conversation.
TLDR
A post contrasts ‘we’re friends’ with ‘we’re just friends,’ arguing that ‘just’ can change how a conversation continues and that LLMs pick up on the distinction. Its author says they show this through human-LLM comparisons, generation experiments and mechanistic interpretability methods.
Combined views
93
3 Sources, first seen ago
