The authors grouped prompt-specific rubric criteria into broader semantic dimensions.
Balancing hidden quality dimensions reduces reward hacking in rubric-trained models
The authors separate aggregate rubric scores into semantic dimensions, then rebalance or complete those dimensions using two reinforcement-learning methods.
Chinese Tech
Jianqiang Xia · Han Xu · Dianxi Shi
Shanghai Jiao Tong University · Baidu, Inc. · Intelligent Game and Decision Lab
Research Digest··3 min read
Xia, Xu and Shi show that rising rubric rewards can conceal declines in specific aspects of response quality, a pattern they call intra-rubric reward divergence.
§