LBR: Towards Mitigating Length Bias in Large Language Models for Recommendation

Jul 16, 2026

Upload by Violet

Large language models (LLMs) increasingly power recommender systems by reframing recommendation as token-level generation, yet a pervasive issue limits their effectiveness: Length Bias. Longer textual item descriptions can dominate input-side attention mass in user modeling, while decoding with summed autoregressive log-likelihood inherently penalizes long outputs. Conventional length normalization may worsen performance. LBR (Length Bias Reduction) provides a lightweight, model-agnostic solution via attention calibration and information-theoretic length normalization, improving accuracy and fairness on multiple Amazon datasets and LLM recommenders.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

Abstract

Introduction

LLM-based recommendation paradigm

Token-level generation and decoding

Length bias motivation

LBR: Towards Mitigating Length Bias in Large Language Models for Recommendation arXiv :2607 .04270v 1 [ cs .IR] 5 Jul 2026 Hongchen Li†‡ [void_jack@zju.edu.cn](void_jack@zju.edu.cn)[ ](void_jack@zju.edu.cn)Zhejiang University Hangzhou, Zhejiang, China Weiqin Yang†‡ [tinysnow@zju.edu.cn](tinysnow@zju.edu.cn)[ ](tinysnow@zju.edu.cn)Zhejiang University Hangzhou, Zhejiang, China Bohao Wang†‡ [bohao.wang@zju.edu.cn](bohao.wang@zju.edu.cn)[ ](bohao.wang@zju.edu.cn)Zhejiang University Hangzhou, Zhejiang, China Hang Pan [hungpaan@mail.ustc.edu.cn](hungpaan@mail.ustc.edu.cn)[ ](hungpaan@mail.ustc.edu.cn)University of Science and Technology of China Hefei, Anhui, China Jingbang Chen [chenjb@cuhk.edu.cn](chenjb@cuhk.edu.cn)[ ](chenjb@cuhk.edu.cn)The Chinese University of Hong Kong Hong Kong, China Bingde Hu [tonyhu@zju.edu.cn](tonyhu@zju.edu.cn)[ ](tonyhu@zju.edu.cn)Bangsun Technology Hangzhou, China Can Wang†‡ [wcan@zju.edu.cn](wcan@zju.edu.cn)[ ](wcan@zju.edu.cn)Zhejiang University Hangzhou, Zhejiang, China Jiawei Chen∗†‡ [sleepyhunt@zju.edu.cn](sleepyhunt@zju.edu.cn)[ ](sleepyhunt@zju.edu.cn)Zhejiang University Hangzhou, Zhejiang, China Abstract Large language models (LLMs) have recently emerged as powerful backbones for recommender systems by reformulating recommendation as a token-level generation task. Despite their promise, we identify a pervasive yet underexplored issue: Length Bias. Because items are represented by textual descriptions of varying lengths, LLM-based recommenders can be systematically biased in two ways. On the input side, longer item descriptions occupy more tokens in the context and thus receive disproportionately large aggregate attention mass during user preference modeling. On the output side, decoding based on summed autoregressive log-likelihood score inherently disfavors long items. Worse still, conventional length normalization can introduce an additional bias and even degrade recommendation performance. To address this problem, we propose LBR (Length Bias Reduction), a lightweight and model-agnostic framework for mitigating length bias in LLM-based recommendation. LBR mitigates input-side bias via Length-Aware Attention Calibration, which incorporates a lengthdependent offset into attention logits to neutralize attention skew. For the output side, LBR introduces Effective Information Length Normalization, replacing naive token count with an informationtheoretic length surrogate derived from the branching structure of ∗ Corresponding author. †State Key Laboratory of Blockchain and Data Security, Zhejiang University.‡College of Computer Science and Technology, Zhejiang University. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission [and/or a fee. Request permissions from permissions@acm.org](and/or a fee. Request permissions from permissions@acm.org). Preprint, Hangzhou, China © 2026 Copyright held by the owner/author(s) . Publication rights licensed to ACM. ACM ISBN XXX-X-XXXX-XXXX-X/XXXX/XX [https://doi.org/XXXXXXX.XXXXXXX](https://doi.org/XXXXXXX.XXXXXXX) the prefix tree. Extensive experiments on three real-world Amazon datasets and two representative LLM-based recommenders demonstrate that LBR substantially alleviates length bias while consistently improving recommendation accuracy and fairness, with negligible additional training and inference overhead (with an average NDCG@5 gain of 16. 82%) . The code is available at [https://github.com/Void-JackLee/LBR](https://github.com/Void-JackLee/LBR).

FAQs

What is Length Bias in LLM-based recommendation?

Length Bias arises because item textual length affects both input-side attention accumulation and output-side score computation during token-level generation and decoding.

How does LBR mitigate input-side length bias?

LBR uses Length-Aware Attention Calibration, adding a length-dependent offset to attention logits to neutralize attention skew caused by longer item descriptions.

How does LBR mitigate output-side length bias?

LBR introduces Effective Information Length Normalization, replacing naive token-count normalization with an information-theoretic length surrogate derived from the prefix tree’s branching structure.

Jul 16, 2026

Upload by Violet