CMU proposes TailRL: RL that optimizes tail rewards, not means

Ruslan Salakhutdinov's team proposes TailRL, which optimizes the tail likelihood of rewards rather than the mean, preserving rare high-quality outputs; it works even with poor initial policy coverage and cuts compute by 256x via GUI sample selection.

2026-09-07 ~ 2026-09-08 · 4 related posts

• TailRL optimizes reward tails instead of mean, cutting samples 256x on GUI grounding — burny_tech · 2026-09-07
• TailRL: New RL Objective Maximizes Upper-Tail Reward Coverage Instead of Just the Mean — burny_tech · 2026-09-07
• Tail-Likelihood RL paper targets upper-tail rewards that mean-based RL masks — burkov · 2026-09-07
• TailRL works even with poor initial policy coverage, authors show in navigation tasks — YouJiacheng · 2026-09-08