Paper Detail

Hybrid Policy Distillation for LLMs

Wenhong Zhu, Ruobing Xie, Rui Wang, Pengfei Liu

huggingface Score 9.5

Published 2026-04-22 · First seen 2026-04-24

General AI

Abstract

Knowledge distillation (KD) is a powerful paradigm for compressing large language models (LLMs), whose effectiveness depends on intertwined choices of divergence direction, optimization strategy, and data regime. We break down the design of existing KD methods and present a unified view that establishes connections between them, reformulating KD as a reweighted log-likelihood objective at the token level. We further propose Hybrid Policy Distillation (HPD), which integrates the complementary advantages of forward and reverse KL to balance mode coverage and mode-seeking, and combines off-policy data with lightweight, approximate on-policy sampling. We validate HPD on long-generation math reasoning as well as short-generation dialogue and code tasks, demonstrating improved optimization stability, computational efficiency, and final performance across diverse model families and scales. The code related to this work is available at https://github.com/zwhong714/Hybrid-Policy-Distillation.

Workflow Status

Review status
pending
Role
unreviewed
Read priority
soon
Vote
Not set.
Saved
no
Collections
Not filed yet.
Next action
Not filled yet.

Reading Brief

No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.

Why It Surfaced

No ranking explanation is available yet.

Tags

No tags.

BibTeX

@misc{zhu2026hybrid,
  title = {Hybrid Policy Distillation for LLMs},
  author = {Wenhong Zhu and Ruobing Xie and Rui Wang and Pengfei Liu},
  year = {2026},
  abstract = {Knowledge distillation (KD) is a powerful paradigm for compressing large language models (LLMs), whose effectiveness depends on intertwined choices of divergence direction, optimization strategy, and data regime. We break down the design of existing KD methods and present a unified view that establishes connections between them, reformulating KD as a reweighted log-likelihood objective at the token level. We further propose Hybrid Policy Distillation (HPD), which integrates the complementary adv},
  url = {https://huggingface.co/papers/2604.20244},
  keywords = {knowledge distillation, large language models, divergence direction, optimization strategy, data regime, reweighted log-likelihood, forward KL, reverse KL, mode coverage, mode-seeking, off-policy data, on-policy sampling, code available, huggingface daily},
  eprint = {2604.20244},
  archiveprefix = {arXiv},
}

Metadata

{}