Search papers, labs, and topics across Lattice.
NVIDIA
1
0
4
12
Speculative decoding, typically used post-RL, can be integrated directly into RL training loops to accelerate LLM rollout generation by up to 2.5x.
