Metraipolis › archive › dc_0e1efb0bf1eb

Sequential Beats Joint: On the Interplay between On-Policy Distillation and RLVR — arXiv:2609.04108

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.04108
License (as archived)
CC0-1.0
Archived
2026-10-17T01:12:20.541Z
SHA-256 of the archived text
4ce1e982d51413bc85585c0cf3971d24264e3e12a8562bc5475b46aa66f9cc42
Stable URL of this page
https://metraipolis.com/documento/dc_0e1efb0bf1eb

How to cite

Sequential Beats Joint: On the Interplay between On-Policy Distillation and RLVR — arXiv:2609.04108 — arXiv (en), https://arxiv.org/abs/2609.04108 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_0e1efb0bf1eb (Metraipolis archive dc_0e1efb0bf1eb, sha256 4ce1e982d51413bc85585c0cf3971d24264e3e12a8562bc5475b46aa66f9cc42)
BibTeX
@misc{dc0e1efb0bf1eb,
  title = {Sequential Beats Joint: On the Interplay between On-Policy Distillation and RLVR — arXiv:2609.04108},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.04108},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_0e1efb0bf1eb; sha256 4ce1e982d51413bc85585c0cf3971d24264e3e12a8562bc5475b46aa66f9cc42},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.04108. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Reinforcement learning with verifiable rewards (RLVR) and on-policy distillation (OPD) have emerged as two dominant methods for post-training reasoning LLMs. Prior work uses OPD's dense token-level supervision to complement the sparse RL reward, fusing the two signals within a single step: either as a \emph{weighted-additive combination} or a \emph{teacher-modulated rescaling} of the RL advantage. In this paper, we show that a simple two-stage scheme, OPD-then-RL, consistently outperforms pure OPD, pure RLVR, and all such joint baselines across logic and math reasoning benchmarks. Beyond the empirical results, we further provide a systematic understanding of this through pass@$k$ behavior, learning dynamics, and parameter updates, yielding a consistent explanation: OPD expands the student's coverage of teacher-supported solutions and RL sharpens within that support, while jointly optimizing the two signals causes them to interfere. To provide a practical recipe, we find that the OPD validation score is the key signal for when to switch to RL, and that OPD is a better cold start for RL than SFT. Together, our results establish OPD-then-RL as a simple yet strong way to combine the two methods, turning two entangled signals into complementary stages.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-04 in cs.AI).
Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI); Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2609.04108 — 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 2026-10-17T00:28:03.035Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-09-04?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).