Get Started
Home
Topics
Search
Library
Research questionHow can RLVR reduce the cost of on-policy rollouts and reliable targets without hurting reasoning quality?RLVR depends on expensive on-policy samples and reliable supervision, making reasoning improvements costly at scale. Efficiency techniques can also be difficult to compare, reuse, and integrate with distributed training.
AI
LLM Pretraining & Post-training
Machine Learning
Reasoning
Reinforcement Learning
Latest papersRecent research connected to this question, newest first.Gradients Know What Outcomes Don't: Unlocking Reinforcement Learning for LLM Reasoning with Gradient-Aligned RewardsApplies to LLM reasoning trained with reinforcement learning from verifiable rewards. The source studies a gradient-based reward using expert anchors on Qwen3-4B and Qwen3-8B, with results on competition-level mathematics, GPQA Diamond, and MMLU-Pro; it reports less than 9% wall-clock overhead and transfer without domain-specific data.research paper · Sep 3, 2026DE-Venus: A Data-Efficient RLVR Framework for Large Language ModelsThe source presents a framework that represents supervision decisions across data preparation and policy optimization, supporting seven representative methods within distributed RL execution. Reported evidence covers public benchmarks and three business scenarios, including configurations using 10% of labels or as little as 13% of relevant data, with selected business cases reducing observed convergence steps by 63%–75%.research paper · Sep 3, 2026
Related questions
Should reasoning LLMs combine on-policy distillation and verifiable-reward RL jointly or sequentially during post-training?How can on-policy diffusion RL reuse rollout computations without trading redundant recomputation for excessive memory use?How can online reinforcement learning make VLA manipulation precise without value drift or prohibitive cost?How can reinforcement learning post-training prioritize useful reasoning prompts as learning signals shift?