1
When Entropy Is Not Enough: Reclaiming Lost Semantics in LLM Output Length Prediction
熵不再是万能指标，这项研究用语义恢复破解LLM输出长度预测难题。
arXiv:2608.15592v1 Announce Type: new Abstract: Efficient LLM serving is often bottlenecked by the need to pad sequences to a fixed maximum length, an…