Open-MOPD addresses the capability imbalance problem in multi-teacher on-policy distillation (M-OPD) by isolating capability integration from routing ambiguity and revealing a 35.6% headroom gap compared to a domain-routed oracle ensemble. The study identifies three key factors—sequence-length disparities, convergence drift, and reward staleness—that misallocate token-level optimization budgets, leading to severe degradation in concise tasks. The proposed Open-MOPD framework introduces token-share balancing, gap-aware dynamic budget allocation, and student reward refresh, boosting headroom recovery to 83.4% and providing an open-source, reproducible post‑training recipe and evaluation suite.
arXiv:2606. 01476v1 Announce Type: new Abstract: On-Policy Distillation (OPD) trains a student model on its own generative trajectories under dense token-level feedback from a stronger teacher, mitigating both the off-policy distribution shift of Supervised Fine-Tuning (SFT) and the sparse credit assignment of Reinforcement Learning (RL).
arXiv:2609.24141v1 Announce Type: new Abstract: On-policy distillation (OPD) pays twice for each fresh batch: the student generates trajectories and a stronger teacher scores them. Existing methods i...
Offline on-policy distillation gains efficiency by collecting student trajectories and teacher supervision once and reusing them throughout optimization. The same reuse makes imperfect supervision per...
arXiv:2605. 07804v3 Announce Type: replace-cross Abstract: On-policy distillation (OPD) leverages dense teacher rewards to enhance reasoning models.
The paper introduces Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation (Road MOPD), a method that adds non-routed teachers as auxiliary cross-domain critics to address gradient conflicts in multi-teacher distillation. It employs token-level credit assignment based on teacher agreement and preventive gradient projection to coordinate updates and reduce conflicts. Experiments show that Road MOPD improves performance, especially in high-conflict scenarios, and mitigates degradation in disadvantaged domains.
Instella‑MoE is a fully open Mixture‑of‑Experts language model with 16 billion total parameters and 2.8 billion active parameters per token, trained from scratch on AMD Instinct GPUs. It incorporates a sparsely activated MoE design with Gated Multi‑head Latent Attention and FarSkip‑Collective connectivity, and follows a multi‑stage pipeline that includes pre‑training, long‑context extension, supervised fine‑tuning, direct preference optimization, and reinforcement learning with Multi‑Teacher On‑Policy Distillation. The model achieves an average score of 76.7 on pre‑training benchmarks and 73.2 on instruction‑following, reasoning, math, coding, and chat benchmarks, outperforming comparable fully open and open‑weight models, and its full training pipeline, weights, and code are released for reproducibility.
arXiv:2608. 03092v1 Announce Type: cross Abstract: We aim to improve model performance in multi-reward reinforcement learning training process.
arXiv:2605. 03677v2 Announce Type: replace Abstract: On-policy distillation (OPD) has recently emerged as an effective post-training paradigm for consolidating the capabilities of specialized expert models into a single student model.
The paper introduces a method for offline on‑policy distillation that addresses the problem of imperfect teacher supervision. By training on teacher‑successful problems and measuring changes in token likelihoods on teacher‑failed trajectories, the authors derive a learnability signal that weights the distillation loss. This approach improves performance on mathematical reasoning and code generation tasks while reducing computational cost compared to online distillation.
arXiv:2607. 13399v1 Announce Type: cross Abstract: On-policy distillation (OPD) has become a key paradigm in LLM post-training, yet its training dynamics remain poorly understood.
arXiv:2607. 16955v1 Announce Type: cross Abstract: On-policy knowledge distillation transfers reasoning from large teachers to compact students, but existing approaches suffer three compounding failure modes: (i) cold-start collapse, where a fresh student assigns near-zero mass to teacher-preferred tokens; (ii) state-agnostic divergence scheduling, where time-only forward/reverse-KL interpolation ignores the student's coverage state; and (iii) binary reward sparsity, where pass/fail signals discard information from partially correct traces.
The paper introduces DIAL-OPD, a token‑selection strategy for on‑policy distillation that weights reward magnitude by the logarithmic mean of teacher and student probabilities, controlled by a parameter β. Experiments on four teacher‑student pairs across seven mathematical reasoning benchmarks show that retaining only 40 % of tokens with DIAL‑OPD yields up to 5.25 % higher mean accuracy than vanilla OPD and doubles AIME25 Pass@16, outperforming even stronger full‑token baselines. The study demonstrates that careful supervision allocation can surpass gains from simply scaling teacher size.