Qiang Wu

Reward hacking can be effectively mitigated by a novel agentic reward framework that enhances exploration in RL, leading to substantial accuracy gains in embodied world models.