@xeophon this would be pretty surprising, considering everything else we know about language models. would you like to try to back this up with a real claim?
I don’t think eval awareness is a real thing // is as much of a problem as people make it out to be
No Digg Deeper questions have been answered for this story yet.
@xeophon @jdchawla29 Oh, if this is the motte then you're really just claiming a lack of evidence in which case why don't you do an experiment improve eval awareness is jumped up. because everybody will be really excited to find this out
@jdchawla29 I’m not doubting that models know what evals, grading etc. is, as its knowledge from the training data
I’m saying that the phenomenon is overstated and its impact is not really assessed // ppl should do more realistic setups
@deepfates @jdchawla29 Want to work on reward hacking on the near future; will comb through the rollouts then to see whether models specifically mention eval settings and if so, how those rollouts differ from others