Uploaded September 2026 | Updated November 2026, 1 day ago
In asynchronous RL with disaggregated training and inference, even slight differences in model behavior between the training and inference stack can cause significant training instabilities that are painful to debug.
At Ray Summit 2026, Dominic Yurk, AI Researcher at Lila Sciences, walks through two contributions upstreamed to SkyRL that close that gap: faithful replay of expert-router decisions and of per-token sample support, both captured at rollout and reconstructed exactly on the training side. He digs into the engineering: a consistent abstraction for attaching token-level metadata, efficient encoding and transport of large-scale router replay data, and a sparse representation that makes sample-support computation fast at scale.
You'll leave with empirical results showing the effect on training stability and throughput.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
In asynchronous RL with disaggregated training and inference, even slight differences in model behavior between the training and inference stack can cause significant training instabilities that are painful to debug.
At Ray Summit 2026, Dominic Yurk, AI Researcher at Lila Sciences, walks through two contributions upstreamed to SkyRL that close that gap: faithful replay of expert-router decisions and of per-token sample support, both captured at rollout and reconstructed exactly on the training side. He digs into the engineering: a consistent abstraction for attaching token-level metadata, efficient encoding and transport of large-scale router replay data, and a sparse representation that makes sample-support computation fast at scale.
You'll leave with empirical results showing the effect on training stability and throughput.
Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo
Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale
đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com










