Search papers, labs, and topics across Lattice.
Affiliation:
1
0
3
6
Asynchronous LLM training can be made robust to stale rollouts by minimizing a novel group-mass capping bias bound, outperforming existing methods on reasoning tasks.
