Metraipolis › archive › dc_e573fdf02dca

When KL Regularization Misfires in Group Policy Optimization — arXiv:2610.12161

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.12161
License (as archived)
CC0-1.0
Archived
2026-10-09T06:13:50.232Z
SHA-256 of the archived text
4f2f9e8e58afee952a1142a53a9f9d7e70ad6a401474a10be3c2811dffd28efe
Stable URL of this page
https://metraipolis.com/documento/dc_e573fdf02dca

How to cite

When KL Regularization Misfires in Group Policy Optimization — arXiv:2610.12161 — arXiv (en), https://arxiv.org/abs/2610.12161 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_e573fdf02dca (Metraipolis archive dc_e573fdf02dca, sha256 4f2f9e8e58afee952a1142a53a9f9d7e70ad6a401474a10be3c2811dffd28efe)
BibTeX
@misc{dce573fdf02dca,
  title = {When KL Regularization Misfires in Group Policy Optimization — arXiv:2610.12161},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.12161},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_e573fdf02dca; sha256 4f2f9e8e58afee952a1142a53a9f9d7e70ad6a401474a10be3c2811dffd28efe},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.12161. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Why does removing reference-policy KL regularization sometimes improve group policy optimization? This motivates studying how reference-policy information should enter group-relative updates. We analyze seven potential failure modes in the interactions between KL and rewards: residual KL updates after reward clipping, after gradient cancellation, and in groups with identical rewards; KL growth with response length and an imbalance in its relative contribution; KL concentration on a small number of tokens; and sampling noise when k1 is incorporated into rewards. We propose Zero-Sum Calibrated Policy Optimization (ZCPO), which uses relative drift measured by conditional KL to calibrate within-group reward coefficients and integrates them into the base surrogate. Mathematical reasoning experiments and ablations support this design's effectiveness in our settings.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.CL, nuovi lavori del giorno).
Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2610.12161 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-10-09T05:47:15.507Z dalla pagina https://arxiv.org/list/cs.CL/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).