arXiv Machine Learning By Ruize Zheng, Xuyang Liu, Siyu Ren, Xiaobing Tu, Yinggui Wang, Jinkui Ren, Xiantao Zhang, Linfeng Zhang

Beyond Passive Imitation: On-Policy Mid-Training Distillation for Language Models

Read the original on arXiv Machine Learning →

The paper introduces On-Policy Mid-Training Distillation (OPMD), a method that shifts continual pre‑training from passive next‑token prediction to active, query‑driven knowledge acquisition. By sampling trajectories from the evolving student policy and applying dense token‑level reverse‑KL rewards from a frozen teacher, OPMD trains the model on states it actually visits, reducing the mismatch between training and generation. Experiments on MATH‑500 and Minerva show that OPMD improves a Qwen student’s performance over its base and outperforms the standard next‑token prediction baseline, while integrating seamlessly into existing continual pre‑training pipelines without inference‑time overhead.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

Hugging Face Trending Papers
Nov 10

Beyond Passive Imitation: On-Policy Mid-Training Distillation for Language Models

The paper introduces On-Policy Mid-Training Distillation (OPMD), a method that replaces the standard next-token prediction (NTP) objective in continual pre‑training with an active, query‑driven approach. By sampling trajectories from the evolving student policy and applying dense token‑level reverse‑KL rewards from a frozen teacher, OPMD trains the model on states it actually visits, reducing the mismatch between training and generation. Experiments on a 1B‑token mathematical corpus show that OPMD improves performance for the Qwen3‑1.7B‑Base model on MATH‑500 and Minerva benchmarks, outperforming the NTP baseline that degrades on both tasks, while adding no inference‑time overhead.

arXiv Machine Learning
Oct 2

Activation-Conditioned Self-Distillation

Activation-Conditioned Self-Distillation (ACSD) is a new on‑policy self‑distillation method that uses a frozen copy of the base model to extract a steering vector by contrasting activations from self‑generated trajectories that reach verified correct answers with all other trajectories. The student learns from next‑token distributions on its own prefixes, without needing reference text or teacher parameter updates, and is used alone at inference. Across five models, ACSD achieves the highest mean accuracy on four mathematical benchmarks, with notable gains on DeepSeek‑R1‑0528‑Qwen3‑8B and LiveCodeBench v6 compared to the OPSD baseline.

By Zhexi Lu, Subhajit Chaudhury, Tejaswini Pedapati, Keerthiram Murugesan, Lei Yu