Tag
Qwen's new paper studies reward design for long-horizon coding agents, showing that every verification signal eventually stops tracking correctness due to reward hacking, and argues verification must co-evolve with policy capability.