Adaptive Supervised Anchoring for On-Policy Self-Distillation
On-policy self-distillation (OPSD) adapts a language model by distilling guidance from a frozen teacher on trajectories sampled from the student. Its effectiveness, however, depends critically on the quality of those trajectories. We show that when student rollouts drift from target trajectories, conditioning the teacher on off-target prefixes substantially weakens its task-relevant supervision. Controlled prefix-corruption experiments expose this failure mode, which we term rollout-conditioned
Record details
Published: 8 August 2026
Source: arXiv cs.LG
Category: Research
Topics: Regulation · Children & education
Retrieved: 11 August 2026
Related evidence
These records share source-supplied organisations, an exact publisher byline, automatic topics or regions. The reason is shown on every link; related does not mean supporting, agreeing with or verifying this record.
Ethical AI use is contextual: focus group insights on students’ decision-making in academic work
AI & Society · 8 August 2026
Where Does AI Innovation Go? Measuring Research Attention Imbalance in AI Music
arXiv · 7 August 2026
Decolonizing Linguistic Policies in Automated Speech Recognition: A Framework for Cross-Culturally Competent Speech AI
arXiv cs.CY · 7 August 2026
Teaching Intro AI When the Tools Can Do the Homework: A Course Redesign and a Student Bill of Rights
arXiv cs.CY · 7 August 2026
Implementation of Split Deadlines in a Large CS1 Course
arXiv · 7 August 2026
Learning from Consensus and Disagreement: Unsupervised On-Policy Self-Distillation with Minority-Trajectory Contrast
arXiv cs.LG · 9 August 2026
How to cite this record
ethics.ai (8 August 2026), “Adaptive Supervised Anchoring for On-Policy Self-Distillation,” evidence record 18336, https://ethics.ai/record/18336 (originally published by arXiv cs.LG).
Use and limitations
This page is a stable index and citation surface for a source record. ethics.ai did not author the underlying report and has not independently verified every claim. Automatic topics may be imperfect. For consequential use, quote and cite the original publisher.