The authors distinguish two broad reward failures.
Adaptive training harness keeps rubric rewards useful as policies improve
HarnessRL jointly adjusts task sampling, rollout guidance, and evaluation criteria based on measured reward failures during training.
Chinese Tech
Xiaoqiang Wang · Ting-En Lin · Yuchuan Wu · Pinyi Zhang · Caixin Kang · Ling Li · +4 more
Alibaba Token Hub, Alibaba Group · Université de Montréal · Mila - Quebec AI Institute
Research Digest··2 min read
Wang et al.
§