Reinforcement learning is now the main lever for turning a strong base model into a capable agent. I work on the full stack of that process: training recipes for reasoning, planning, tool use, and agentic coding; reward modeling and rubric-based rewards, including their failure modes such as reward hacking (Yang et al., 2026); RL recipes that resist reward hacking and efficient multi-agent systems; and the asynchronous, high-throughput RL systems that make long-horizon agent training practical. RSI Arena (Huang et al., 2026) takes this toward recursive self-improvement with human feedback: ten frontier models act as training agents, each post-training the same base model on a shared cluster with a fixed compute and API budget, and then refining it from blind human comparisons of its own outputs.
@article{huang2026rsiarena,title={{RSI} Arena: Towards Recursive Self-Improvement with Human Feedback},author={Huang, Yue and Wang, Yanbo and Li, Lichi and Chen, Yuanyuan and Haupt, Andreas and Luo, Xiaonan and Liu, Zheyuan and Choi, Jihye and Juefei-Xu, Felix and Zhang, Xiangliang and Poovendran, Radha and Sra, Misha and Liu, Ying and Burtenshaw, Ben and Wolf, Thomas and Pentland, Alex and Chen, Zichen and Xu, Zhangchen},journal={arXiv preprint arXiv:2610.33624},year={2026},month=oct,}
@article{yang2026rubric,title={Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward {RL}},author={Yang, Minglai and Guo, Xinyu and Tyagi, Utkarsh and Zhang, Mian and Dumitru, Razvan and Hou, Sunjie and He, Yunzhong and Zhang, Daniel Yue and Liu, Ying},journal={arXiv preprint arXiv:2608.11669},year={2026},month=aug,}