Confidence is constant in adaptive sampling

[akashgupta222]

I noticed that the adaptive sampling technique uses a confidence measure for each token which is added to the log prob. However, the inputs to the confidence score as the same for all sampled tokens. This makes confidence score constant across all tokens, effectively leading the argmax only on the log probs. If this is the intention, should the confidence calculation be removed from the codebase to make it easier to understand ? If not, how should the confidence measure be adapted to make it token dependent ?

entropix/entropix/sampler.py

Line 300 in 9d41be6

confidence_score = (