Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Zhang, Junjie, Yang, Rushuai, Liu, Shunyu, Lin, Ting-En, Huang, Fei, Chen, Yi, Li, Yongbin, Tao, Dacheng
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2504.07527
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866918040911740928
author Zhang, Junjie
Yang, Rushuai
Liu, Shunyu
Lin, Ting-En
Huang, Fei
Chen, Yi
Li, Yongbin
Tao, Dacheng
author_facet Zhang, Junjie
Yang, Rushuai
Liu, Shunyu
Lin, Ting-En
Huang, Fei
Chen, Yi
Li, Yongbin
Tao, Dacheng
contents In this work, we establish a novel theoretical connection between supervised fine-tuning and offline reinforcement learning under the token-level Markov decision process, revealing that large language models indeed learn an implicit $Q$-function for inference. Through this theoretical lens, we demonstrate that the widely used beam search method suffers from unacceptable over-optimism, where inference errors are inevitably amplified due to inflated $Q$-value estimations of suboptimal steps. To address this limitation, we propose Supervised Optimism Correction(SOC), which introduces a simple yet effective auxiliary loss for token-level $Q$-value estimations during supervised fine-tuning. Specifically, the auxiliary loss employs implicit value regularization to boost model confidence in expert-demonstrated responses, thereby suppressing over-optimism toward insufficiently supervised responses. Extensive experiments on mathematical reasoning benchmarks, including GSM8K, MATH, and GAOKAO, showcase the superiority of the proposed SOC with beam search across a series of open-source models.
format Preprint
id arxiv_https___arxiv_org_abs_2504_07527
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Supervised Optimism Correction: Be Confident When LLMs Are Sure
Zhang, Junjie
Yang, Rushuai
Liu, Shunyu
Lin, Ting-En
Huang, Fei
Chen, Yi
Li, Yongbin
Tao, Dacheng
Computation and Language
In this work, we establish a novel theoretical connection between supervised fine-tuning and offline reinforcement learning under the token-level Markov decision process, revealing that large language models indeed learn an implicit $Q$-function for inference. Through this theoretical lens, we demonstrate that the widely used beam search method suffers from unacceptable over-optimism, where inference errors are inevitably amplified due to inflated $Q$-value estimations of suboptimal steps. To address this limitation, we propose Supervised Optimism Correction(SOC), which introduces a simple yet effective auxiliary loss for token-level $Q$-value estimations during supervised fine-tuning. Specifically, the auxiliary loss employs implicit value regularization to boost model confidence in expert-demonstrated responses, thereby suppressing over-optimism toward insufficiently supervised responses. Extensive experiments on mathematical reasoning benchmarks, including GSM8K, MATH, and GAOKAO, showcase the superiority of the proposed SOC with beam search across a series of open-source models.
title Supervised Optimism Correction: Be Confident When LLMs Are Sure
topic Computation and Language
url https://arxiv.org/abs/2504.07527