Decagon describes RL methods for a hundreds-of-billions-parameter customer-facing agent

Decagon's Cyrus Asgari discusses reinforcement learning for a production customer-facing agent with hundreds of billions of parameters, according to a post from LMSYS Org on X. Asgari covers synthetic data, reward design, and stabilizing reinforcement learning under training-inference mismatch.

The post presents these methods as ways to build better production agents. It gives no model release, benchmark results, or performance figures.