analiza aiqbench
This research addresses "privilege-induced style drift" in on-policy self-distillation, a phenomenon where alignment signals inadvertently favor stylistic shortcuts over task-relevant reasoning. The proposed RLCSD method utilizes contrastive learning between correct and incorrect hints to suppress these style shifts, resulting in more robust training for mathematical and logical models. By concentrating supervision on task-bearing tokens rather than stylistic artifacts, this approach offers a significant improvement over existing GRPO and prior OPSD techniques for advanced reasoning systems.
RLCSD: Reinforcement Learning with Contrastive On-Policy Self-Distillation
arXiv:2606.11709v2 Announce Type: replace-cross Abstract: On-policy self-distillation (OPSD) provides dense, token-level supervision for reasoning models by aligning a model's own distribution with that under privileged context, typically a verified solution. However, we show that the resulting distributional gap concentrates on style tokens rather than task-bearing ones, as the hinted model tends to produce shorter, more direct outputs. We term this pathology \emph{privilege-induced style drift}, which can destabilize training and shorten responses. To address this, we propose \textbf{RLCSD} (Reinforcement Learning with Contrastive on-policy Self-Distillation), which mitigates this drift by contrasting the teacher-student gap under a correct hint against that under a wrong hint, suppressing style shifts induced by hints regardless of correctness and yielding a signal more concentrated on task-bearing tokens. Experiments on Qwen3 (1.7B/4B/8B) and Olmo-3-7B-Think across mathematical and logical reasoning show that RLCSD consistently outperforms GRPO and prior OPSD methods, with additional results on agentic tasks supporting broader applicability. We further show that the contrastive principle is general: it plugs into existing OPSD methods to improve them, and its underlying insight extends to broader cross-model on-policy distillation.
Abstract:On-policy self-distillation (OPSD) provides dense, token-level supervision for reasoning models by aligning a model's own distribution with that under privileged context, typically a verified solution. However, we show that the resulting distributional gap concentrates on style tokens rather than task-bearing ones, as the hinted model tends to produce shorter, more direct outputs. We term this pathology \emph{privilege-induced style drift}, which can destabilize training and shorten responses. To address this, we propose \textbf{RLCSD} (Reinforcement Learning with Contrastive on-policy Self-Distillation), which mitigates this drift by contrasting the teacher-student gap under a correct hint against that under a wrong hint, suppressing style shifts induced by hints regardless of correctness and yielding a signal more concentrated on task-bearing tokens. Experiments on Qwen3 (1.7B/4B/8B) and Olmo-3-7B-Think across mathematical and logical reasoning show that RLCSD consistently outperforms GRPO and prior OPSD methods, with additional results on agentic tasks supporting broader applicability. We further show that the contrastive principle is general: it plugs into existing OPSD methods to improve them, and its underlying insight extends to broader cross-model on-policy distillation.
Machine Learning (cs.LG); Computation and Language (cs.CL)
arXiv:2606.11709 [cs.LG]
(or arXiv:2606.11709v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.11709 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 269 cuvinte
Subiecte conexe
RLCSD: Reinforcement Learning with Contrastive On-Policy Self-Distillation
arXiv · 12 iun. 2026
Contrastive Reinforced Policy Optimization via Privileged Self-Distillation
arXiv · 31 iul. 2026
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
SIPO: Unifying Reinforcement Learning with On-Policy Self-Distillation
arXiv · 1 oct. 2026