Introduction of ILRec, a two‑stage preference‑fine‑tuning framework for LLM‑based recommendation.
Extraction of self‑hard negative tokens from intermediate layers to provide fine‑grained, dynamic negative supervision.
Cross‑layer preference optimization and distillation jointly improve discriminative capability and the quality of negative signals.
A lightweight collaborative filtering module assigns token‑level rewards to reduce over‑penalization of false negatives.
Empirical results on three datasets confirm ILRec’s superior performance over existing baselines.
📖 Full Retelling
WHO: The paper is authored by Bingqian Li, Bowen Zheng, Xiaolei Wang, Long Zhang, Jinpeng Wang, Sheng Chen, Wayne Xin Zhao, and Ji‑rong Wen. WHAT: They propose ILRec, a preference‑fine‑tuning framework that uses self‑hard negatives extracted from intermediate layers of large language models to enhance recommendation performance. WHERE: The work was submitted to the arXiv preprint server under the Information Retrieval (cs.IR) and Artificial Intelligence (cs.AI) categories. WHEN: It was first posted on arXiv on 19 February 2026. WHY: The authors identified that existing negative‑sampling strategies are coarse and offline, limiting model discriminativeness in large item spaces, and sought a dynamic, fine‑grained approach to improve LLM‑based recommender systems.
ILRec identifies self‑hard negative tokens from intermediate layers of an LLM to provide granular negative supervision that evolves with the model’s learning process. It employs a two‑stage training scheme—cross‑layer preference optimization followed by cross‑layer preference distillation—to jointly discriminate informative negatives and refine the negative signals. A lightweight collaborative filtering model supplies token‑level rewards, mitigating the risk of penalizing false negatives. Extensive experiments on three benchmark datasets demonstrate that ILRec consistently outperforms prior methods in recommendation accuracy.
🏷️ Themes
Recommendation systems, Large language models, Preference learning, Negative sampling, Intermediate layer representation, Cross‑layer optimization, Collaborative filtering, Fine‑grained supervision
Entity Intersection Graph
No entity connections available yet for this article.
Deep Analysis
Why It Matters
The paper introduces ILRec, a framework that uses self-hard negatives from intermediate layers of LLMs to improve recommendation systems. This addresses limitations of offline-generated negatives and enhances fine-grained preference learning, potentially leading to more accurate and efficient recommender systems.
Context & Background
LLMs are increasingly used in recommendation tasks
Traditional preference learning relies on sequence-level offline negatives
Intermediate layer representations can capture richer negative signals
Cross-layer optimization and distillation are novel techniques
Experiments show improved performance on three datasets
What Happens Next
The ILRec framework may be adopted by industry and academia to build next-generation recommender systems. Future research could explore scaling to larger datasets, integrating with other modalities, and evaluating real-world deployment.
Frequently Asked Questions
What are self-hard negatives?
Tokens from intermediate layers that the model finds difficult, used as fine-grained negative supervision.
How does ILRec differ from existing methods?
It uses dynamic, model-derived negatives instead of static offline negatives and includes cross-layer distillation.
What datasets were used?
Three benchmark recommendation datasets; specific names are not provided in the abstract.
Original Source
--> Computer Science > Information Retrieval arXiv:2602.17410 [Submitted on 19 Feb 2026] Title: Improving LLM-based Recommendation with Self-Hard Negatives from Intermediate Layers Authors: Bingqian Li , Bowen Zheng , Xiaolei Wang , Long Zhang , Jinpeng Wang , Sheng Chen , Wayne Xin Zhao , Ji-rong Wen View a PDF of the paper titled Improving LLM-based Recommendation with Self-Hard Negatives from Intermediate Layers, by Bingqian Li and 7 other authors View PDF HTML Abstract: Large language models have shown great promise in recommender systems, where supervised fine-tuning is commonly used for adaptation. Subsequent studies further introduce preference learning to incorporate negative samples into the training process. However, existing methods rely on sequence-level, offline-generated negatives, making them less discriminative and informative when adapting LLMs to recommendation tasks with large negative item spaces. To address these challenges, we propose ILRec, a novel preference fine-tuning framework for LLM-based recommendation, leveraging self-hard negative signals extracted from intermediate layers to improve preference learning. Specifically, we identify self-hard negative tokens from intermediate layers as fine-grained negative supervision that dynamically reflects the model's preference learning process. To effectively integrate these signals into training, we design a two-stage framework comprising cross-layer preference optimization and cross-layer preference distillation, enabling the model to jointly discriminate informative negatives and enhance the quality of negative signals from intermediate layers. In addition, we introduce a lightweight collaborative filtering model to assign token-level rewards for negative signals, mitigating the risk of over-penalizing false negatives. Extensive experiments on three datasets demonstrate ILRec's effectiveness in enhancing the performance of LLM-based recommender systems. Subjects: Information Retrieval (cs.IR) ;...
