#OPSD
A study introduces Termination-Aware On-Policy Self-Distillation (TA-OPSD), optimizing when language models should continue or stop. TA-OPSD reduces premature terminations and repetitions, improving performance on math reasoning tasks. https://arxiv.org/abs/2611.19925
Learning When to Continue and When to Stop: Termination-Aware OPSD
ArXiv link for Learning When to Continue and When to Stop: Termination-Aware OPSD
arxiv.org
November 18, 2026 at 6:20 PM
View of landing at Skardu Airport (OPSD) located in the Himalayan mountain range, Pakistan.

Original post
November 10, 2026 at 2:05 PM
👩‍🏫OPSD(On-policy Self-distillation)では参照情報付きの教師を固定するか応答単位で変えるのみだったが、好ましい教師はトークンごとに異なる
👩‍🏫そこで複数教師から多数決・KL基準で各トークンが教師を選ぶMTA-OPSDを提案し、Qwen3の数学推論でベース比最大+5.9改善した
arxiv.org/abs/2610.40574
Every Token Needs Its Own Teacher: Multi-teacher Adaptive On-policy Self-distillation
On-policy Self-distillation (OPSD) has proven effective for post-training large language models by constructing a teacher conditioned on privileged references. However, existing approaches keep the te...
arxiv.org
November 6, 2026 at 1:24 AM
Qingyuan Zheng, Zhihao Dou: Shuffled OPSD: When Semantic Consistency Hurts On-Policy Self-Distillation? https://arxiv.org/abs/2611.04958 https://arxiv.org/pdf/2611.04958 https://arxiv.org/html/2611.04958
November 4, 2026 at 6:40 AM
Search-Aware On-Policy Self-Distillation with Sibling Alignment for Generative Recommendation

Distills self-teacher knowledge at beam-search prefixes a student visits, focusing on those near pruning and keeping teacher rankings of sibling branches.

📝 arxiv.org/abs/2611.01568
Search-Aware On-Policy Self-Distillation with Sibling Alignment for Generative Recommendation
Generative recommendation retrieves items by autoregressively generating semantic IDs, typically via beam search. On-policy self-distillation (OPSD) enhances training supervision by transferring privi...
arxiv.org
November 4, 2026 at 6:52 AM
Kaixuan Du, Xin Wan, Dai Guan, Ming Chen, Xiaoxi Jiang, Guanjun Jiang: Live-OPSD: On-Policy Self-Distillation for Proactive Streaming Large Video Models https://arxiv.org/abs/2610.36959 https://arxiv.org/pdf/2610.36959 https://arxiv.org/html/2610.36959
October 28, 2026 at 6:43 AM
A new chapter for Duke’s Strong Start Award Program! @sudarrajagopal.bsky.social has been named Director, helping support the next generation of physician-scientists. Read more in the October OPSD Newsletter: t.e2ma.net/webview/wyni...

#PhysicianScientist #ResearchCareers #Mentorship #DukeMed
October 27, 2026 at 3:08 PM
Not Every Rollout Needs Distillation: Correctness-Filtered On-Policy Self-Distillation

Harrison Zhao et al.

#arXiv #cs.AI
Not Every Rollout Needs Distillation: Correctness-Filtered On-Policy Self-Distillation
On-policy self-distillation (OPSD) supervises student-generated trajectories with a reference-conditioned version of the same base model. We ask whether every trajectory supplies distinct supervision that warrants its distillation cost. Gradient measurements on Qwen3-1.7B at initialization reveal t…
arxiv.org
October 20, 2026 at 6:22 PM
Researchers have introduced E2-OPSD, an innovative method for on-policy self-distillation that addresses "entropy overshoot" in AI training. This approach significantly enhances math reasoning without requiring additional networks. https://arxiv.org/abs/2610.05048
E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
ArXiv link for E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
arxiv.org
October 17, 2026 at 4:10 AM
Mitigating Capability Degradation in LLM Self-Improvement via Representation Geometry Regularization

Jiaqi Tian, Ting Long

#arXiv #cs.AI
Mitigating Capability Degradation in LLM Self-Improvement via Representation Geometry Regularization
On-Policy Self-distillation is a self-improvement paradigm that enables large language models (LLMs) to improve their capabilities by leveraging the same model as both teacher and student under different conditioning contexts. However, we find that OPSD methods can progressively degrade previously …
arxiv.org
October 12, 2026 at 9:40 PM
Mitigating Capability Degradation in LLM Self-Improvement via Representation Geometry Regularization

Jiaqi Tian, Ting Long

#arXiv #cs.AI
Mitigating Capability Degradation in LLM Self-Improvement via Representation Geometry Regularization
On-Policy Self-distillation is a self-improvement paradigm that enables large language models (LLMs) to improve their capabilities by leveraging the same model as both teacher and student under different conditioning contexts. However, we find that OPSD methods can progressively degrade previously …
arxiv.org
October 12, 2026 at 8:15 PM
On-Policy Self-distillation enables large language models to improve capabilities using the same model as teacher and student. However, OPSD methods can progressively degrade previously acquired capabilities during optimization.

Source: arXiv
Mitigating Capability Degradation in LLM Self-Improvement via Representation Geometry Regularization
Abstract page for arXiv paper 2610.12855: Mitigating Capability Degradation in LLM Self-Improvement via Representation Geometry Regularization
arxiv.org
October 12, 2026 at 8:30 AM
A study on E2-OPSD tackles "entropy overshoot" in self-distillation, improving math reasoning by up to 4.9 points. It employs exemplar-guided teaching and entropy-aware distillation to boost learning efficiency without extra models, setting a new standard in AI. https://arxiv.org/abs/2610.05048
E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
ArXiv link for E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
arxiv.org
October 9, 2026 at 3:00 PM
E2-OPSD enhances on-policy self-distillation by addressing "entropy overshoot," boosting math reasoning performance by up to 4.3 points. It maintains simplicity, requiring no extra models or computations, thus improving training efficiency. https://arxiv.org/abs/2610.05048
E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
ArXiv link for E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
arxiv.org
October 9, 2026 at 8:20 AM
A study introduces E2-OPSD, a method for self-distillation that reduces entropy overshoot in language models, enhancing math reasoning by 4.3 points. It employs exemplar-guided teaching and entropy-aware distillation, optimizing AI learning without extra models. https://arxiv.org/abs/2610.05048
E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
ArXiv link for E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
arxiv.org
October 7, 2026 at 6:00 PM
A study presents E2-OPSD, a method for on-policy self-distillation that tackles "entropy overshoot" in LLM training, boosting math reasoning by 4.3 points. With exemplar-guided teaching and entropy-aware distillation, it ensures stable learning without extra models. https://arxiv.org/abs/2610.05048
E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
ArXiv link for E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
arxiv.org
October 7, 2026 at 5:50 PM
E2-OPSD tackles "entropy overshoot" in on-policy self-distillation, boosting math reasoning by 4.3 points. It employs exemplar-guided teaching and entropy-aware distillation, stabilizing learning and reducing dependence on specific answer cues. https://arxiv.org/abs/2610.05048
E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
ArXiv link for E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation
arxiv.org
October 7, 2026 at 1:30 PM
Yifei Liu, Minghao Fang, Xinyu Gu, Chengkai Yao, Mengdi Liu, Tengfei Ma, Jiangbin Zheng, Chang Yu, Zhangyang Gao: E$^2$-OPSD: Taming Entropy Overshoot in On-Policy Self-Distillation https://arxiv.org/abs/2610.05048 https://arxiv.org/pdf/2610.05048 https://arxiv.org/html/2610.05048
October 6, 2026 at 6:44 AM
2/n💡Our idea:

Give the teacher a textual hint telling it which visual elements are relevant and where they are located in the image.

Through OPSD, the student learns from this privileged teacher to locate and integrate evidence from multiple relevant image regions.
October 5, 2026 at 1:23 PM
Deng, Xing, Xu, He, Zheng, Li, Zhu, Li: Can Vision-Language Models Stay Helpful When Facing Implicit Risks? Intent-Privilege OPSD for Efficient Safety-Helpfulness Alignment https://arxiv.org/abs/2609.37837 https://arxiv.org/pdf/2609.37837 https://arxiv.org/html/2609.37837
October 1, 2026 at 6:41 AM
Train Ahead, Distill Back: Bootstrapping On-Policy Self-Distillation for Large Language Models

Zheng Zhang et al.

#arXiv #cs.AI
Train Ahead, Distill Back: Bootstrapping On-Policy Self-Distillation for Large Language Models
On-policy self-distillation (OPSD) improves large language models by letting a self-teacher with privileged information provide dense token-level supervision on the model's own trajectories. Yet existing methods typically construct the self-teacher from the current, initial, or slowly averaged poli…
arxiv.org
September 30, 2026 at 8:20 PM
Train Ahead, Distill Back: Bootstrapping On-Policy Self-Distillation for Large Language Models

Zheng Zhang et al.

#arXiv #cs.AI
Train Ahead, Distill Back: Bootstrapping On-Policy Self-Distillation for Large Language Models
On-policy self-distillation (OPSD) improves large language models by letting a self-teacher with privileged information provide dense token-level supervision on the model's own trajectories. Yet existing methods typically construct the self-teacher from the current, initial, or slowly averaged poli…
arxiv.org
September 30, 2026 at 6:46 PM
Zhenyu Liu, Zhangquan Chen, Keyi Chen, Mingze Sun, Xiang An, Haodong Jing, Ruqi Huang: Spatial-OPSD: Self-Improving Spatial Reasoning via Label-Free Self-Distillation https://arxiv.org/abs/2609.37055 https://arxiv.org/pdf/2609.37055 https://arxiv.org/html/2609.37055
September 30, 2026 at 6:41 AM
Learning from a Thoughtful Teacher: Adaptive On-Policy Self-Distillation for Mathematical Reasoning

Jiacheng Du et al.

#arXiv #cs.AI
Learning from a Thoughtful Teacher: Adaptive On-Policy Self-Distillation for Mathematical Reasoning
On-policy self-distillation (OPSD) trains a question-only student with token-level feedback from a teacher given training-only privileged information (PI). OPSD therefore provides dense, on-policy supervision, and is free of a larger external teacher, but its effectiveness rests on how PI is design…
arxiv.org
September 29, 2026 at 11:45 PM