AI & ML
Which Tokens, in What Order, and How Fast? A Theoretical Analysis of On-Policy Distillation Objectives
arXiv stat.ML0 views
arXiv:2611.23397v1 Announce Type: cross Abstract: On-policy distillation (OPD) trains a student language model to match a teacher model based on its own generations. When using OPD, one must choose whether to apply it at the sequence or token level and which divergence to minimize. However, despite the wide adoption of OPD in practice, the formal understanding of which OPD objective is appropriate for a given setting is limited. Toward addressing this gap, we theoretically analyze how the choice of OPD objective governs which tokens are prioritized, in what order, and how efficiently. Our analysis reveals that sequence-level OPD exhibits a future agreement bias, which myopic token-level objectives avoid. This bias induces a reverse learning order, in which sequence-level OPD first matches the teacher at later token positions, while at earlier positions it can favor tokens that are improbable under the teacher, and thus often undesirable. As for the choice of divergence, we prove that it strongly influences the amenability of OPD to gradient-based optimization. In particular, while the forward KL divergence enjoys a well-behaved objective landscape, other $f$-divergences, such as reverse KL, can suffer from vanishing gradients even when the student and teacher are far apart. Across language model post-training settings, we demonstrate the implications of these theoretical findings for which OPD objective is preferable, highlighting the importance of accounting for the interplay between the student, teacher, and task at hand.
Source:https://arxiv.org/abs/2611.23397PreviousExploration via Exploitation: The Blessing of Reward Diversity in Personalized Federated RLNextWhen Does Subspace Direction Matter for LoRA? Regime Analysis of the Magnitude Principle in Few-Shot Adaptation
Related posts
Designing a useful preview for an AI-generated comic
0dev.to
AI Agents: What They Are and Why They Matter
0dev.to
GPT-6.2 Astra Was Easier to Jailbreak, UK Testing Found
0dev.to
iFixit Says Apple’s HomeView Breaks Easily. Here’s Why I’m Not Worried
0CNET News
Scout Motors Is Using AI Earpieces to Train Factory Workers
0CNET News
'Grand Theft Auto VI' is here. What to know about the characters, music and all the hype
0Phys.org Tech