Saved in:
Bibliographic Details
Main Authors: Fan, Jiabin, Luo, Guoqing, Bowling, Michael, Mou, Lili
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2504.19024
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916709001068544
author Fan, Jiabin
Luo, Guoqing
Bowling, Michael
Mou, Lili
author_facet Fan, Jiabin
Luo, Guoqing
Bowling, Michael
Mou, Lili
contents We propose a novel k-step return estimation method (called KETCHUP) for Reinforcement Learning(RL)-based knowledge distillation (KD) in text generation tasks. Our idea is to induce a K-step return by using the Bellman Optimality Equation for multiple steps. Theoretical analysis shows that this K-step formulation reduces the variance of the gradient estimates, thus leading to improved RL optimization especially when the student model size is large. Empirical evaluation on three text generation tasks demonstrates that our approach yields superior performance in both standard task metrics and large language model (LLM)-based evaluation. These results suggest that our K-step return induction offers a promising direction for enhancing RL-based KD in LLM research.
format Preprint
id arxiv_https___arxiv_org_abs_2504_19024
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle KETCHUP: K-Step Return Estimation for Sequential Knowledge Distillation
Fan, Jiabin
Luo, Guoqing
Bowling, Michael
Mou, Lili
Computation and Language
We propose a novel k-step return estimation method (called KETCHUP) for Reinforcement Learning(RL)-based knowledge distillation (KD) in text generation tasks. Our idea is to induce a K-step return by using the Bellman Optimality Equation for multiple steps. Theoretical analysis shows that this K-step formulation reduces the variance of the gradient estimates, thus leading to improved RL optimization especially when the student model size is large. Empirical evaluation on three text generation tasks demonstrates that our approach yields superior performance in both standard task metrics and large language model (LLM)-based evaluation. These results suggest that our K-step return induction offers a promising direction for enhancing RL-based KD in LLM research.
title KETCHUP: K-Step Return Estimation for Sequential Knowledge Distillation
topic Computation and Language
url https://arxiv.org/abs/2504.19024