Reinforcement Learning from Verifiable Rewards works well when a task has a programmatic checker, but most long-horizon agent domains have none. We work in t…
机构:CMU
来源:arXiv 2609.04094 | AI4Papers 论文推荐平台