Do models know when they are being evaluated?
Today's Topic
This week Giles Edkins will present his research on whether language models know when they are being evaluated.
One of the concerns with evaluating the alignment and capabilities of models is that they will detect that they are in an evaluation setting and modify their behaviour to appear less dangerous. At present little is known about what distinguishes real user conversations from automated evaluations and whether the language model is wise to these differences.
Giles has been researching what we can learn about this (together with Joe Needham and Govind Pimpale, mentored by Marius Hobbhahn as part of the MATS program).
Read more on Less Wrong here.
We welcome a variety of backgrounds, opinions and experience levels.
Event Schedule
6:00 to 6:45 - Networking and refreshments
6:45 to 8:00 - Main Presentation
Is there a topic you'd love to see us cover at a future event? Submit your suggestion here.