Search papers, labs, and topics across Lattice.
This paper identifies hard clipping in GRPO-style RLVR objectives as a bottleneck that discards informative signals near the clipping boundary, leading to training instability and suboptimal convergence. They propose Near-boundary Stochastic Rescue (NSR), a method that stochastically retains tokens slightly beyond the clipping threshold to recover these lost signals. Experiments across 7B-30B models show NSR improves training stability and performance compared to DAPO and GSPO, demonstrating the importance of near-boundary signals.
Rigid reward clipping throws away valuable information just beyond the boundary, but a simple stochastic rescue of these signals can substantially boost RLVR performance.
Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a central paradigm for scaling LLM reasoning, yet its optimization often suffers from training instability and suboptimal convergence. Through a systematic dissection of clipping-based GRPO-style objectives, we identify the rigid clipping decision induced by hard clipping as a key practical bottleneck in the studied RLVR setups. Specifically, our analysis suggests that informative signals can lie in the near-boundary region just beyond the clipping threshold, and are therefore discarded by the standard hard-clipping rule. Notably, once this bottleneck is precisely identified, even simple stochastic perturbations at the boundary can recover meaningful performance gains. Building on this finding, we propose Near-boundary Stochastic Rescue (NSR), a minimal, plug-and-play modification that stochastically retains these slightly out-of-bound tokens to recover lost signals. While NSR, via stochastic sampling, can be interpreted as inducing an implicit gradient decay in expectation, our ablations reveal that its stochastic, boundary-local rescue mechanism is consistently more effective than deterministic gradient decay. Validated by extensive experiments across model sizes from 7B to 30B and both dense and MoE architectures, as a plug-and-play solution, NSR substantially improves training stability and delivers consistent gains over strong baselines such as DAPO and GSPO.
