Sign up to save your podcasts
Or
Share Not All Rollouts are Useful: Down-Sampling Rollouts in LLM Reinforcement Learning
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
Not All Rollouts are Useful: Down-Sampling Rollouts in LLM Reinforcement Learning
PODS decouples reinforcement learning phases by parallelizing rollouts and selectively updating, using max-variance down-sampling to enhance performance on the GSM8K benchmark compared to standard GRPO.
https://arxiv.org/abs//2504.13818
YouTube: https://www.youtube.com/@ArxivPapers
TikTok: https://www.tiktok.com/@arxiv_papers
Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016
Spotify: https://podcasters.spotify.com/pod/show/arxiv-papers
View all episodes
By Igor Melnyk
5
33 ratings
PODS decouples reinforcement learning phases by parallelizing rollouts and selectively updating, using max-variance down-sampling to enhance performance on the GSM8K benchmark compared to standard GRPO.
https://arxiv.org/abs//2504.13818
YouTube: https://www.youtube.com/@ArxivPapers
TikTok: https://www.tiktok.com/@arxiv_papers
Apple Podcasts: https://podcasts.apple.com/us/podcast/arxiv-papers/id1692476016
Spotify: https://podcasters.spotify.com/pod/show/arxiv-papers
More shows like Arxiv Papers
View all
Exchanges
979 Listeners
Odd Lots
2,003 Listeners
The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence)
436 Listeners
The Daily
113,168 Listeners
All-In with Chamath, Jason, Sacks & Friedberg
10,270 Listeners
Hard Fork
5,542 Listeners
UnHerd with Freddie Sayers
218 Listeners
Unsupervised Learning with Jacob Effron
54 Listeners
Latent Space: The AI Engineer Podcast
100 Listeners
BG2Pod with Brad Gerstner and Bill Gurley
460 Listeners