LLM latency is high enough that per-call gradient steps and per-batch threshold solves can fit inside the round-trip time

At LLM latency, per-call gradient steps and per-batch threshold solves fit inside the round-trip.

http://arxiv.org/abs/2608.27244v1