5arXiv cs.CL (Computation and Language)·
On-Policy Delta Distillation improves reasoning transfer via teacher-base model difference signal
Researchers from NAVER AI introduce On-Policy Delta Distillation (OPD²), a new post-training method that replaces direct imitation of a teacher model's output distribution with a 'delta signal' — the difference between the teacher and its pre-instruction-tuning base model. This delta signal isolates the reasoning capability changes induced by instruction tuning, providing a more targeted supervision signal for student models. Experiments across math, science, and code-reasoning benchmarks show OPD² consistently outperforms conventional on-policy distillation with shorter post-training periods.
Related guides (2)
Related events (8)
Direct On-Policy Distillation transfers RL policy shifts from weak to strong models
Researchers propose Direct-OPD (Direct On-Policy Distillation), a method for transferring the policy shift induced by reinforcement learning on a small model to a larger target model, bypassing the need to run expensive RL rollouts on the stronger model. The approach uses the log-ratio between a post-RL teacher and its pre-RL reference as a dense implicit reward signal applied to the student's own on-policy states. Empirically, Direct-OPD improves Qwen3-1.7B from 48.3% to 62.4% on AIME 2024 in 4 hours on 8 A100 GPUs, outperforming step-matched direct RL. The method addresses a key scalability bottleneck in post-training as frontier models grow larger.
Purified OPSD fixes on-policy self-distillation failures in long chain-of-thought reasoning models
A new arXiv preprint identifies why on-policy self-distillation (OPSD) consistently degrades long chain-of-thought reasoning models: the teacher's supervision signal is dominated by reference-induced shortcuts rather than question-conditioned, transferable corrections. The authors propose a two-step fix using a reference-only teacher to isolate the non-transferable component and pointwise mutual information (PMI) to construct a cleaner distillation target. Experiments across four long-CoT models on two datasets show consistent improvements over both the base model and standard OPSD while preserving reflective reasoning behavior.
Are Full Rollouts Necessary for On-Policy Distillation?
This paper investigates whether full rollouts are required during on-policy distillation (OPD) for training reasoning models, identifying rollout horizon as a key computational bottleneck. The authors propose two strategies: Progressive OPD (POPD), which gradually expands rollout horizon during training, and Truncated OPD (TOPD), which uses permanently truncated rollouts. Experiments on mathematical reasoning show POPD achieves up to 3× training efficiency improvement, while TOPD matches full OPD performance using only 10% of the rollout horizon, yielding significant wall-clock and memory savings.
Analysis of on-policy distillation reveals sparse, geometrically structured parameter updates
A new arXiv paper analyzes on-policy distillation (OPD) — a post-training method combining on-policy student trajectories with dense teacher supervision — across language and vision-language model pairs. The authors find that OPD updates are coordinate-sparse and distributed across layers (FFN-heavy), and that training only the discovered sparse subnetwork recovers near-full performance. Geometrically, updates are numerically full-rank but spectrally concentrated, falling disproportionately on near-zero weight coordinates, suggesting OPD retains distinct geometric signatures rather than behaving like ordinary dense parameter rewriting.
OPCoD: On-Policy Co-Distillation for Mutual LLM Improvement via Peer Feedback
Researchers introduce On-Policy Co-Distillation (OPCoD), a training framework where two LLMs, each stronger in a different domain, iteratively tutor each other using on-policy rollouts and peer feedback. The method uses cognizance-based gating to control when feedback is given and feedback anchoring to ground it in the problem context. On Science Q&A tasks, OPCoD achieves Pareto improvement for both models across all evaluated domain pairs, outperforming one-way distillation and single-model fine-tuning baselines.
MOPD: Multi-Teacher On-Policy Distillation for integrating multiple RL-trained capabilities in LLMs
Researchers propose Multi-teacher On-Policy Distillation (MOPD), a post-training paradigm that first trains domain-specialized RL teacher models, then distills them into a student model using on-policy rollouts to eliminate exposure bias. Evaluated on Qwen3-30B-A3B, MOPD outperforms Mix-RL, Cascade RL, Off-Policy Finetune, and Param-Merge baselines while preserving nearly all per-domain capability. The method has been deployed in production for MiMo-V2-Flash, an industrial-scale frontier model, validating its practical applicability. The approach also enables parallel, decoupled development of domain teachers, reducing cross-domain interference in multi-capability post-training.
DOPD: Advantage-aware dual on-policy distillation to address privilege illusion in LLM/VLM training
Researchers introduce DOPD (Dual On-policy Distillation), a knowledge distillation framework that dynamically routes token-level supervision between a privileged teacher and privileged student policy based on advantage gap and relative probabilities. The method addresses a failure mode called 'privilege illusion,' where information asymmetry between teacher and student is conflated with a transferable capability gap. Experiments on both LLM and VLM settings show DOPD outperforms vanilla on-policy distillation and related methods, with additional gains on stability, continual learning, and out-of-distribution tasks.
DemoPSD: Disagreement-modulated policy self-distillation to fix privileged information leakage in LLM reasoning training
DemoPSD is a new training framework for LLMs that addresses two failure modes in on-policy self-distillation (OPSD): overfitting to in-domain patterns and privileged information leakage, where the student model learns answer-dependent shortcuts unavailable at test time. The method steers the student toward a reverse-KL barycenter target — a weighted geometric blend of teacher and student distributions — with token-level blending weights derived from the disagreement between the two distributions. Experiments on SciKnowEval across four scientific domains show DemoPSD outperforms GRPO and SDPO while maintaining higher training entropy and generalizing to out-of-distribution GPQA benchmarks.