Opposite scaling law: detection of machine-generated text is done better by smaller models
Everyone (outside #NLProc...) is afraid GPT would cheat for them, which pushes for detection methods
https://arxiv.org/abs/2305.09859
#NLProc #ML #machinelearning
First the problem, given a text you want to know whether a human wrote it. You've been in NLP lately I am sure a teacher, sister, nephew etc. called and told you they suspect someone handed them a GPT text.
Problem: how can you tell
The approach
Randomly replace words
Then see how much it changed the sentence probability\likelihood
presented by
https://arxiv.org/abs/2301.11305
The idea behind it
Model text is uniquely fit to the model expectations
Human texts are just arbitrary good sentences
Therefore, any change to a machine text would have a larger effect on the probability
What this paper found is that
the smaller the LM you pick the better it is in detecting the LLM text
regardless of same training data same architecture, model etc.
As always a small model behaves just like an undertrained model
https://twitter.com/LChoshen/status/1506245430912430091
https://arxiv.org/abs/2109.06096
So you can also pick an undertrained model, which would be better than a trained model in detecting who generated the text.