NudgeRL: Structured Exploration for Reinforcement Learning with Verifiable Rewards
A new framework called NudgeRL proposes structured and diversity-driven exploration for reinforcement learning with verifiable rewards (RLVR) in large language models. The approach introduces Strategy Nudging, which conditions rollouts on lightweight strategy-level contexts to generate diverse reasoning trajectories without expensive oracle supervision. A unified objective decomposes the reward signal to improve learning efficiency. The work addresses the fundamental limitation of RLVR where policy improvement is constrained by previously sampled trajectories, offering an alternative to computationally expensive brute-force scaling. The paper is available on arXiv under identifier 2605.15726.
Key facts
- NudgeRL is a framework for structured exploration in RLVR
- Strategy Nudging conditions rollouts on strategy-level contexts
- A unified objective decomposes the reward signal
- RLVR improves reasoning capabilities of large language models
- Exploration is limited by previously sampled trajectories
- Brute-force scaling is computationally expensive
- The paper is on arXiv with ID 2605.15726
Entities
Institutions
- arXiv