LEEPS: Latent-Guided Explore-Exploit Prompt Sampling for Efficient RLVR in Large Language Models

LEEPS is a latent-guided pre-rollout prompt sampler for RLVR that dynamically splits prompts into exploit and explore groups to gather more useful learning signals.

It ranks uncertain prompts using latent-space neighbors and past outcomes, reducing wasted rollouts on redundant generations.

On Qwen2.5-Math-1.5B and 7B, it delivered the best average results across six math reasoning benchmarks and three out-of-distribution reasoning benchmarks.

The method adds only modest training overhead while improving both training efficiency and reasoning performance in RLVR.