Nan Jiang is a member of the technical staff at Modal building infrastructure for cross-datacenter reinforcement learning and contributing to the open-source post-training frameworks slime and miles. His work makes geographically scattered GPU capacity available for training advanced models without requiring every machine to join the same high-speed cluster.
Jiang studied computer science at the University of California, San Diego, investigating whether attention is necessary for in-context learning. As a research assistant working with Alexander Rush at Cornell Tech, he helped develop Commit0, which challenges coding agents to implement complete software libraries, and GameArena, which evaluates model reasoning through interactive games. He also coauthored RouterBench, a framework for comparing language-model routing strategies against performance and cost. Before joining Modal, he worked at Amazon’s AGI San Francisco lab on computer-use agents, reinforcement-learning post-training, multimodal models, and training infrastructure.
- Separating trainers from rollout fleets. Jiang keeps backpropagation and collective operations on RDMA-connected GPUs while distributing trajectory generation across independent serving clusters. Each cluster receives a policy version and returns trajectories with associated metadata, allowing rollout capacity to expand across regions and providers.
- Lossless served-weight synchronization. Shipping complete model checkpoints between datacenters can make distributed training prohibitively slow. Jiang instead synchronizes changed parameter positions and replacement values in the serving engine’s numerical format, reconstructing identical weights without accumulating approximation errors. His merged contribution to slime implements disaggregated rollout synchronization through disk and NCCL transports.
- Adam absorption and finite precision. Small Adam updates can change high-precision master weights without crossing the rounding thresholds of BF16, FP8, or other serving formats. The resulting sparsity appears between successive served-weight versions, even when gradients and optimizer updates remain dense.
- Version-aware elastic rollout fleets. Modal’s Stitch combines immutable policy versions, explicit readiness checks, and workers that report which model version generated each trajectory. This architecture turns geographically dispersed inference capacity into an elastic reinforcement-learning resource while preserving synchronization and staleness guarantees.
Jiang is exploring whether similar synchronization efficiencies hold for Muon, how asynchronous reinforcement learning scales, and whether distributed rollouts can extend to supervised fine-tuning and other training stages.
