arXiv:2610.27450v1 Announce Type: new Abstract: On-policy distillation (OPD) can succeed with surprisingly few training questions, motivating a view of questions as carriers of distributional knowled...
The paper investigates how much training data is required for on‑policy distillation (OPD) of foundation models. It shows that in two single‑teacher settings, only eight real prompts can match the performance of 17,000 problems, and that 64 self‑generated questions can replace full external datasets. In multi‑teacher scenarios, 1,000 generated questions achieve comparable results to 7,000 real examples, while fully implicit question generation works only in limited cases.
The paper proposes Mixture Proposal On-Policy Distillation (MPOPD) to address a mismatch issue in on‑policy distillation (OPD) where the rollout policy is fixed to the current student. By treating the rollout policy as an optimization target, MPOPD derives a geometric mixture of student and teacher policies and introduces a proposal policy for efficient rollout generation. Experiments on 16 reasoning benchmarks show consistent gains over OPD, notably a 6.91% accuracy improvement on math reasoning with Qwen3-4B-Base and a 41% reduction in generation length.
arXiv:2607. 02502v1 Announce Type: cross Abstract: On-policy self-distillation (OPSD) has emerged as a practical method for training large language models (LLMs) to reason, where a single model acts as both the teacher and the student with different levels of information access.
arXiv:2609.14193v1 Announce Type: cross Abstract: On-policy distillation (OPD) has become a standard component of frontier post-training pipelines, yet how much its training data actually contributes...
The paper introduces a recursive self-improvement framework for language models that replaces an external teacher with a frozen copy of the student, enabling dynamic co-evolution (DCE) and self-refined concise learning (SRCL). DCE allows the privileged teacher to evolve alongside the student, while SRCL trains on shorter, verified rewrites to reduce verbosity. Experiments show that the combined DCE+SRCL approach outperforms traditional on‑policy self‑distillation across multiple model sizes and math benchmarks, achieving significant accuracy gains and shorter outputs.
arXiv:2609.38025v1 Announce Type: cross Abstract: On-policy distillation (OPD) trains a student on its own generated responses using dense, token-level supervision from a stronger teacher. Vanilla OP...
On-policy distillation (OPD) trains a student on its own generated responses using dense, token-level supervision from a stronger teacher. Vanilla OPD treats all teacher signals equally, assuming that...
arXiv:2607. 02234v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) has emerged as a promising paradigm for improving LLM reasoning, where a privileged teacher with access to reference solutions provides token-level supervision on the student's own generated trajectories.
Post-training of reasoning language models is commonly driven by supervised distillation and reinforcement learning with verifiable rewards. Distillation often relies on chain-of-thought annotations that are expensive to obtain and may themselves be noisy, incomplete, or partially incorrect; even when the final solution is correct, an imperfect rationale can interfere with learning.
The paper introduces Monte Carlo Privileged Distillation (MCPD) to address reference-induced target instability in on-policy self-distillation (OPSD). MCPD models privileged context as a distribution over partial reference views and averages teacher distributions across these views, yielding a target that matches the expected KL divergence minimization. Experiments on mathematical and scientific reasoning benchmarks show that MCPD provides more reference-consistent supervision and consistently outperforms full-reference OPSD.
On-policy Distillation (OPD) supervises a student model on trajectories sampled from its own policy by minimizing the divergence between the output distributions of the teacher and student at each token position, thereby providing dense token-level supervision. Although existing OPD methods have demonstrated strong performance in improving the reasoning ability of student models, their objectives fundamentally rely on token-level distribution matching.