Including temperature scaling makes forward KL divergence outperform reverse KL in LLM distillation on instruction benchmarks, overturning the τ=1 preference for reverse KL.
arXiv preprint arXiv:2212.12965 , year=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Rethinking the Role of Temperature in Large Language Model Distillation
Including temperature scaling makes forward KL divergence outperform reverse KL in LLM distillation on instruction benchmarks, overturning the τ=1 preference for reverse KL.