RARE addresses open-ended tasks where outputs can be valid in multiple ways and overall quality lacks an automatic verifier.
Adaptive rubrics improve reinforcement learning for open-ended text generation
RARE selectively replaces saturated evaluation criteria, with flaw-focused updates producing the strongest results across video scripts and other writing tasks.
Chinese Tech
Xueyun Tian · Bingbing Xu · Chuyue Li · Jiaming Liu · Chao Zhang · Weijia Feng · +6 more
State Key Laboratory of AI Safety, Institute of Computing Technology, CAS · University of Chinese Academy of Sciences · Qwen Business Unit of Alibaba · University of Illinois at Urbana-Champaign
Research Digest··3 min read
Tian and colleagues developed RARE, a reinforcement-learning method that updates rubric criteria as a model’s capabilities change.
§