Study: Noisy data severely degrades RLVR training performance

ddkang · x · 2026-08-28

Research by Daniel Kang's team refutes the claim that RLVR is robust to noisy data. They found prior claims of success with '100% incorrect' data were due to contamination with correct labels. A rigorous curation pipeline shows noisy data degrades test accuracy by over 9%, which algorithmic improvements fail to mitigate.

Related event: Cleaned BIRD-Platinum Dataset Propels Kimi-K2.6 Past Frontier Models in Text-to-SQL(3 posts)→

Original post →

More from Research

Research channel →