Follow
Unbelievably simple recent ideas in ML, often top-conference fodder:
"to increase performance by 10% absolute, just take the majority-vote answer of several LM answers"
"to reduce resource use by 50%(!), use a large model to do rejection sampling of small models' output"
"to find hyperparams about twice as fast, start a bunch of networks training and after a while copy the weights of the one improving fastest. repeat"
https://www.deepmind.com/blog/population-based-training-of-neural-networks
I guess chain of thought is itself one of these.
To detect if text comes from LM X, randomly modify it and get X's logprobs of the original and the mod.
If p(original) > p(mod), classify as LM generated.
https://arxiv.org/abs/2301.11305v1