RLM-Cascade: Response-Level Speculative Decoding for Cost-Efficient LLM API Serving

[system]

---

This is a companion discussion topic for the original entry at https://arxiv.org/abs/2606.22840