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

Large Language Models (LLMs) are increasingly used for sequential recommendation by treating item prediction as a token-level generation task. However, the authors identify a pervasive 'Length Bias' that undermines system performance. Because items have varying textual lengths, they are systematically misrepresented in two ways: first, longer item descriptions receive disproportionately high attention mass during user preference modeling; second, standard autoregressive decoding penalizes longer sequences, while naive length normalization fails to account for the varying informativeness of tokens in a constrained prefix tree (Trie) decoding environment.

To mitigate these biases, the authors introduce LBR (Length Bias Reduction), a model-agnostic framework that operates at both the input and output stages:

• Length-Aware Attention Calibration (LAAC): This module introduces a length-dependent offset into the attention logits. By neutralizing the correlation between item length and cumulative attention mass, it ensures that the model's focus on historical items is based on relevance rather than sequence length.
• Effective Information Length Normalization (EILN): Instead of normalizing by raw token count, this approach quantifies token informativeness using Hartley entropy derived from the branching factor of the Trie. This allows the model to normalize scores by the 'information content' of the sequence, providing a fairer comparison between candidates with different lengths and structural complexities.

This research is significant because it addresses a fundamental, often overlooked, structural bias in generative recommendation systems. By demonstrating that length bias is not merely a data artifact but a byproduct of the Transformer architecture and constrained decoding, the authors provide a principled way to improve both recommendation accuracy and fairness. The proposed LBR framework is lightweight, requires negligible computational overhead, and consistently outperforms existing debiasing methods across multiple real-world datasets, making it a highly practical solution for deploying LLMs in production recommender systems.