3 ms·It’s a problem specific to autoregressive LLMs, the early tokens bias the outputby mountainriver 1y agoIt’s a problem specific to autoregressive LLMs, the early tokens bias the output