Paper Detail

Denser $\neq$ Better: Limits of On-Policy Self-Distillation for Continual Post-Training

Meng Wang, Haohan Zhao, Wenzhuo Liu, Lu Yang, Geng Liu, Haiyang Guo, Guo-Sen Xie, Gaofeng Meng, Hongbin Liu, Fei Zhu

arxiv Score 18.8

Published 2026-07-02 · First seen 2026-07-03

Research Track A · General AI

Abstract

Continual post-training enables foundation models to acquire new knowledge while preserving existing capabilities. Recent work suggests that on-policy learning can mitigate forgetting, with on-policy self-distillation emerging as a particularly attractive approach. In this work, we revisit this optimistic view through self-distillation policy optimization (SDPO). Our experiments show that SDPO can accelerate in-domain specialization when teacher signals are stable and well aligned, but it struggles to generalize to out-of-distribution scenarios. In continual post-training, SDPO exhibits stronger forgetting and can even collapse, whereas on-policy reinforcement learning methods such as GRPO adapt more conservatively and better preserve prior capabilities. Further analyses reveal that denser self-distillation induces larger drift in both parameter space and response space, and can amplify high-frequency formatting artifacts through a self-reinforcing teacher--student loop. These findings suggest that on-policy data alone is insufficient for continual learning. Dense self-distillation can accelerate specialization when teacher targets are stable and token-level supervision is reliable, but it should not be treated as a default stabilizer for continual post-training. Our code is available at https://github.com/Moenupa/SDPO-CL.

Workflow Status

Review status
pending
Role
unreviewed
Read priority
now
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

@article{wang2026denser,
  title = {Denser \$\textbackslash{}neq\$ Better: Limits of On-Policy Self-Distillation for Continual Post-Training},
  author = {Meng Wang and Haohan Zhao and Wenzhuo Liu and Lu Yang and Geng Liu and Haiyang Guo and Guo-Sen Xie and Gaofeng Meng and Hongbin Liu and Fei Zhu},
  year = {2026},
  abstract = {Continual post-training enables foundation models to acquire new knowledge while preserving existing capabilities. Recent work suggests that on-policy learning can mitigate forgetting, with on-policy self-distillation emerging as a particularly attractive approach. In this work, we revisit this optimistic view through self-distillation policy optimization (SDPO). Our experiments show that SDPO can accelerate in-domain specialization when teacher signals are stable and well aligned, but it strugg},
  url = {https://arxiv.org/abs/2607.01763},
  keywords = {cs.LG, cs.CL, continual post-training, on-policy learning, self-distillation, on-policy self-distillation, policy optimization, continual learning, model specialization, model forgetting, reinforcement learning, GRPO, parameter space, response space, self-reinforcing loop, code available, huggingface daily},
  eprint = {2607.01763},
  archiveprefix = {arXiv},
}

Metadata

{}