← Back to all news
arXiv Machine Learning By Jingru Yang, Jinhang Wu, Minjing Shi, Shengfeng He, Xuyan Li, Man Tan, Biye Li

CAST: Credit Alignment for Cross-Tokenizer On-Policy Distillation

Read the original on arXiv Machine Learning →

CAST (Chunk-aware credit Assignment with Selective Token Anchoring) is a new framework for cross-tokenizer on-policy distillation that tackles two key problems: credit alignment and unaligned token positions. It redistributes chunk-level advantage signals across student tokens using predictive uncertainty and teacher-signal smoothness, while anchoring unaligned tokens to the initial policy via a KL penalty. Experiments on six mathematics, science, and code benchmarks show that CAST improves average performance over the strongest baseline by up to 2.45 points in cross-tokenizer transfer scenarios.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

Related stories

arXiv Machine Learning
Oct 22

CSD: Common-Support Distillation for Cross-Tokenizer Knowledge Transfer

The paper introduces Common‑Support Distillation (CSD), a logits‑based framework for transferring knowledge across different tokenizers by constructing a shared support between teacher and student distributions. CSD handles both one‑to‑one alignments using explicit common‑token pairs and other alignments by creating a virtual support via the autoregressive chain rule, while a residual event ensures normalization. Experiments on Llama‑3.2‑1B, Gemma3‑1B‑PT, Qwen, and Phi models show that CSD consistently outperforms existing cross‑tokenizer methods, achieving up to 1.8‑point gains over the best alternative and surpassing both continued pretraining and same‑tokenizer distillation.

By Mingyu Yang, Sharath Turuvekere Sreenivas, Adithyakrishna Venkatesh Hanasoge, Nima Tajbakhsh, Ashwath Aithal, Pavlo Molchanov
llmscomputer-visionnlpefficiency
More like this →
arXiv Computation and Language
Oct 16

SimCT: Recovering Lost Supervision for Cross-Tokenizer On-Policy Distillation

SimCT is a method for on‑policy distillation that addresses the problem of mismatched tokenizers between teacher and student models. By comparing teacher and student predictions over short multi‑token continuations that both tokenizers can produce, SimCT restores supervision that would otherwise be lost when only exact shared tokens are used. Experiments on mathematical reasoning and code‑generation tasks show consistent improvements over shared‑vocabulary OPD and other cross‑tokenizer baselines, with ablations confirming the gains come from recovering discarded supervision.

By Jie Sun, Mao Zheng, Mingyang Song, Qiyong Zhong, Yilin Cheng, Bichuan Feng, Pengfei Liu, Junfeng Fang, Xiang Wang
nlpefficiencybenchmarks
More like this →
arXiv AI
Aug 11

Mismatch Matters: On-Policy Distillation Beyond Token Agreement

arXiv:2608. 09836v1 Announce Type: new Abstract: On-policy distillation (OPD) has emerged as a core component of modern LLM post-training pipelines, yet we reveal a failure mode: degenerate agreement, where students exploit repetitive loops to achieve near-perfect token agreement with the teacher despite globally flawed responses.

By Zichao Yu, Chengzhi Yu, Shengze Xu, Yujin Han, Bingqing Jiang, Xu Wang, Difan Zou
llmsefficiencybenchmarks
More like this →