Potential Typo in the Paper

[XuRunhui]

I wonder whether this is a typo in your formula or I have wrongly understood you idea.

[dpaleka]

The equation in the paper is wrong, thanks! But the fix is a bit different.
We add B to both tokens i and R:

stealing-part-lm-supplementary/query_logprobs_emulator/query_logprobs.py

Line 364 in 6fa1d1a

logit_bias[anchor_token_id] = bias

The final equation is y_R^B - y_i^B = z_R - z_i. We will update this in the next version.
The following sentence in the paper might make it clearer:

Since we can observe 5 logprobs, we can compare the reference token R to four tokens per query, by adding a large bias that pushes all four tokens into the top 5 (along with the reference token).