SPIN Unprocessed October 15, 2026 ai_technology research
Token Utility for On-Policy Distillation via Local Trust-Region Optimization
View original on arxiv.orgOverview
arXiv:2610.18883v1 Announce Type: new Abstract: Token-selection methods for on-policy distillation (OPD) prioritize student-visited positions using diverse signals, including uncertainty, teacher-student disagreement, and mismatch direction. We instead derive token importance from the distillation objective itself: how much can each position's loss decrease under an equal local KL budget on the student's predictive distribution? The answer is the utility $s_t=g_t^\top F_t^\dagger g_t$, where $g_
SpinGraph analysis pending — check back after processing.
Ask AI about this story
Opens with the SpinGraph .md URL and structured context — one click, prompt included.
More from arXiv Machine Learning
View all →- Backpropagating Verifier Rewards Through Continuous Flow Language Models
- Theoretical and Practical Advances in Hierarchical Bayesian Deep Learning Methodology
- How Inner Optimizers Shape Replica Scaling
- Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes
- What Information Buys in Active Inference
- Graph Coarsening Enables Escaping Local Optima in Compute-Constrained Nonconvex MAX-CUT
Markdown (.md) · JSON-LD schema (.json) · Machine-readable for AI & GEO