TailRL optimizes reward tails instead of mean, cutting samples 256x on GUI grounding
burny_tech · x · 2026-09-07
Researchers propose Tail-Likelihood RL (TailRL), which targets a flaw in standard RL: optimizing average reward suppresses rare but exceptional rollouts and hurts Best-of-k scaling. TailRL instead maximizes the log-probability of exceeding a randomly chosen reward threshold, weighting rare high-reward samples more.
- Its gradient is a harmonic mixture of Best-of-k gradients across all k
- Requires only a simple advantage-function tweak, compatible with existing RL pipelines
- On GUI grounding it needs up to 256x fewer samples; code optimization Best-of-1024 sees a 7.7x speedup
- Evaluated on object localization, maze navigation, GUI grounding, and code optimization; code released as Zanette-Labs/TailRL
Related event: CMU proposes TailRL: RL that optimizes tail rewards, not means(4 posts)→
More from Research
- MissionReport benchmark: only 3 of 540 AI-generated government reports pass all criteria — Exp_Mark · 2026-10-27
- Finch 1: a geospatial foundation model trained on imagery from 50+ sensors — Scobleizer · 2026-10-27
- Skyfall AI's SKY-001 world model turns sketches into CAD builds in 3.9s, vs 199s for GPT-6 Astra — rohanpaul_ai · 2026-10-27
- Physicists Let AI Pick Its Own Problem—It Wrote a Publishable Paper in 2 Hours — burny_tech · 2026-10-27
- Researcher flags unresolved data transparency issues in compositionality evals — najoungkim · 2026-10-27
- Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD? — johnhewtt · 2026-10-27