Interesting - I wonder if it's because coding doesn't use the specific token probabilities, while decision models do
replyI think it also helps that speed and latency aren't as vital for coding and we can afford to let the model think for longer.
replyYes this is the reason. Transformations like quantisation preserve the rank of outcomes much better than probability mass.
reply