Multiple research papers published on arXiv explore ways to make reinforcement learning (RL) more efficient and effective for large language models.
According to arxiv.org, researchers introduced GRLO (Generalizable Reinforcement Learning in Open-Ended Environments), which aims to reduce the computational burden of RL-based post-training. The paper states that GRLO improves average performance across all domains from 24.1 to 63.1 using a Qwen3-4B-Base backbone “with only 5K prompts and 22.7 GPU hours, requiring about $46\times$ less data and $68\times$ less compute than a strong in-domain RLVR baseline.” The resulting model was described as “competitive with Qwen’s released post-trained models which required a much larger training cost.”
In a separate paper, arxiv.org introduced AstraFlow, a “dataflow-oriented RL system” designed to support complex multi-policy agentic RL workloads. According to the source, AstraFlow “achieves comparable or better accuracy than existing RL systems while speeding up training time by 2.7x” in multi-policy collaborative training across math, code, search, and AgentBench workloads.
Additionally, arxiv.org presented Prefix-RFT, a method that combines supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). The paper, accepted to ICML 2026, demonstrates that this hybrid approach “surpasses the performance of standalone SFT and RFT” and “outperforms parallel mixed-policy RFT methods” when tested on mathematical reasoning problems.