The authors developed CCE-OPD, a modification of on-policy distillation, in which a student learns from teacher feedback on trajectories generated by the student itself.
Condition-aware distillation reduces invalid reasoning in smaller language models
By contrasting teacher predictions with and without a question condition, the method steers student models away from continuations that contradict stated premises.
Chinese Tech
Wenxuan Luo · Chenxi Li · Haoyue Liu · Yongxin Guo · Xiaoying Tang
The Chinese University of Hong Kong, Shenzhen · Shenzhen Future Network of Intelligence Institute (FNii-Shenzhen) · ByteDance · Guangdong Provincial Key Laboratory of Future Networks of Intelligence, CUHK(SZ)
Research Digest··2 min read
Luo et al.
Why this paper
From ByteDance and 3 others
In one line
Condition-contrastive distillation, which compares teacher scores with and without a question condition, reduces invalid continuations and improves distilled LM reasoning.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓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.
§