The authors address a core limitation in rubric-based RL for open-ended tasks: standard methods aggregate criterion-level evaluations into a scalar reward and assign the same advantage to all tokens, losing fine-grained feedback.
Criterion-aware credit assignment improves rubric-based reinforcement learning for language models
CoCA derives signed criterion-level credits from group outcomes and allocates them across tokens using privileged rubric information.
Chinese Tech
Guangcheng Zhu · Yingfan Ma · Xing Zheng · Haobo Wang · Shenzhi Yang · Kai Tang · +4 more
Zhejiang University · Ant Group
Research Digest··3 min read
The authors propose CoCA, a criterion-aware credit assignment framework for rubric-based reinforcement learning (RL) of large language models.
Why this paper
From Ant Group and Zhejiang University
In one line
CoCA improves rubric-based reinforcement learning by 4.2% on medicine and science benchmarks through criterion-aware credit assignment.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§