arXiv:2609.37500v1 Announce Type: new
Abstract: On-policy distillation (OPD) trains language models using dense token-level teacher supervision on student-generated trajectories. However, its relianc...
By Yuxiao Yang, Shangzhe Li, Tianrun Yu, Kaixiang Zhao, Taylor W. Killian, Weitong Zhang
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
arXiv:2604. 05634v2 Announce Type: replace Abstract: Machine unlearning (MU) has become a critical technique for GenAI models' safe and compliant operation.
By Zhiyong Ma, Zhitao Deng, Huan Tang, Jialin Chen, Zhijun Zheng, Zhengping Li, Qingyuan Chuai
arXiv:2609.36695v1 Announce Type: new
Abstract: Self-distillation turns knowledge distillation into a closed learning loop and offers a path toward recursive self-improvement. Without an external tea...
By Rui Wang, Ruijie Wang, Bo Chen, Jiangxuan Long, Yingyu Liang
arXiv:2606. 09456v1 Announce Type: new Abstract: On-Policy Distillation (OPD) has become a core technique in the post-training of Large Language Models (LLMs) for transferring knowledge from domain experts to student models.
By Yifan Niu, Han Xiao, Dongyi Liu, Zelong Wang, Dihong Gong, Yasheng Wang, Jia Li
VISTA is an online self‑distillation framework that enforces consistency along a deep learning model’s optimization trajectory. It uses a validation‑informed Marginal Coverage score to identify earlier model states—called expert anchors—that retain specialized competence over distinct data regions. By integrating a coverage‑weighted ensemble of these anchors during training, VISTA regularizes the loss landscape, preserves learned knowledge, and improves robustness and generalization while cutting storage overhead by 90%.
By Eli Corn, Daphna Weinshall