Unverified50% confidenceFactExact time
RLVR (reinforcement learning with verifiable rewards) performs excellently on in-distribution tasks, but its performance drops significantly (policy collapse) when transferred out of distribution (OOD).
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe optimization landscape of RLVR is fundamentally different from supervised learning: the reward signal is sparse and discrete, gradient estimates come from sampled responses, and the variance is significantly higher than in traditional training settings.80% similarUnverifiedCompared to reinforcement learning from human feedback (RLHF), RLVR does not rely on human annotators, has extremely low marginal cost, and can be run at a scale far exceeding that of RLHF.76% similarUnverifiedThe jev architecture used for comparison employs parallel sampling, is trained via RLCD (Reinforcement Learning from Contrastive Distillation), and outputs confidence distributions and schema selection.73% similarUnverifiedSeveral recent empirical works have questioned the actual utility of first-order momentum in reinforcement learning with verifiable rewards (RLVR).73% similarUnverified强化学习训练比监督学习更不稳定，将NVFP4激进量化引入RL训练会进一步放大固有不稳定性73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130325API
curl https://kongchang.com/api/v1/knowledge/claims/1130325MCP
get_claim(id=1130325)