Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Token-level Collaborative Alignment for LLM-based Generative Recommendation

Fake Lin, Binbin Hu, Zhi Zheng, Xi Zhu, Ziqi Liu, Zhiqiang Zhang, Jun Zhou, Tong Xu

TL;DR

This work tackles the misalignment between item-level collaborative filtering signals and token-level training in LLM-based recommendations. It introduces TCA4Rec, a model-agnostic framework with two components: Collaborative Tokenizer, which maps item-level CF logits into token-aligned distributions, and Soft Label Alignment, which fuses these with one-hot targets to form a soft NTP objective. The approach preserves the generative nature of LLMs while enabling explicit CF-guided optimization, and it demonstrates consistent gains across diverse CF models and LLM backbones, along with a dedicated collaborative consistency metric. The results indicate that explicit token-level CF supervision improves recommendation accuracy and controllability, offering a practical pathway for integrating traditional CF knowledge into generative recommender systems.

Abstract

Large Language Models (LLMs) have demonstrated strong potential for generative recommendation by leveraging rich semantic knowledge. However, existing LLM-based recommender systems struggle to effectively incorporate collaborative filtering (CF) signals, due to a fundamental mismatch between item-level preference modeling in CF and token-level next-token prediction (NTP) optimization in LLMs. Prior approaches typically treat CF as contextual hints or representation bias, and resort to multi-stage training to reduce behavioral semantic space discrepancies, leaving CF unable to explicitly regulate LLM generation. In this work, we propose Token-level Collaborative Alignment for Recommendation (TCA4Rec), a model-agnostic and plug-and-play framework that establishes an explicit optimization-level interface between CF supervision and LLM generation. TCA4Rec consists of (i) Collaborative Tokenizer, which projects raw item-level CF logits into token-level distributions aligned with the LLM token space, and (ii) Soft Label Alignment, which integrates these CF-informed distributions with one-hot supervision to optimize a soft NTP objective. This design preserves the generative nature of LLM training while enabling collaborative alignment with essential user preference of CF models. We highlight TCA4Rec is compatible with arbitrary traditional CF models and generalizes across a wide range of decoder-based LLM recommender architectures. Moreover, it provides an explicit mechanism to balance behavioral alignment and semantic fluency, yielding generative recommendations that are both accurate and controllable. Extensive experiments demonstrate that TCA4Rec consistently improves recommendation performance across a broad spectrum of CF models and LLM-based recommender systems.

Token-level Collaborative Alignment for LLM-based Generative Recommendation

TL;DR

This work tackles the misalignment between item-level collaborative filtering signals and token-level training in LLM-based recommendations. It introduces TCA4Rec, a model-agnostic framework with two components: Collaborative Tokenizer, which maps item-level CF logits into token-aligned distributions, and Soft Label Alignment, which fuses these with one-hot targets to form a soft NTP objective. The approach preserves the generative nature of LLMs while enabling explicit CF-guided optimization, and it demonstrates consistent gains across diverse CF models and LLM backbones, along with a dedicated collaborative consistency metric. The results indicate that explicit token-level CF supervision improves recommendation accuracy and controllability, offering a practical pathway for integrating traditional CF knowledge into generative recommender systems.

Abstract

Large Language Models (LLMs) have demonstrated strong potential for generative recommendation by leveraging rich semantic knowledge. However, existing LLM-based recommender systems struggle to effectively incorporate collaborative filtering (CF) signals, due to a fundamental mismatch between item-level preference modeling in CF and token-level next-token prediction (NTP) optimization in LLMs. Prior approaches typically treat CF as contextual hints or representation bias, and resort to multi-stage training to reduce behavioral semantic space discrepancies, leaving CF unable to explicitly regulate LLM generation. In this work, we propose Token-level Collaborative Alignment for Recommendation (TCA4Rec), a model-agnostic and plug-and-play framework that establishes an explicit optimization-level interface between CF supervision and LLM generation. TCA4Rec consists of (i) Collaborative Tokenizer, which projects raw item-level CF logits into token-level distributions aligned with the LLM token space, and (ii) Soft Label Alignment, which integrates these CF-informed distributions with one-hot supervision to optimize a soft NTP objective. This design preserves the generative nature of LLM training while enabling collaborative alignment with essential user preference of CF models. We highlight TCA4Rec is compatible with arbitrary traditional CF models and generalizes across a wide range of decoder-based LLM recommender architectures. Moreover, it provides an explicit mechanism to balance behavioral alignment and semantic fluency, yielding generative recommendations that are both accurate and controllable. Extensive experiments demonstrate that TCA4Rec consistently improves recommendation performance across a broad spectrum of CF models and LLM-based recommender systems.
Paper Structure (37 sections, 30 equations, 2 figures, 7 tables)

This paper contains 37 sections, 30 equations, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Sequential Recommendation Task
  4. LLM-based Recommendation Task
  5. Methodology
  6. Collaborative Tokenizer
  7. Collaborative Logits Derivation
  8. Token-level Distribution Transformation
  9. Soft Label Alignment
  10. Relation to Auxiliary KL Loss
  11. Auxiliary KL Formulation
  12. Gradient Analysis
  13. Experiment
  14. Settings
  15. Dataset
  16. ...and 22 more sections

Figures (2)

  • Figure 1: The overall framework of our method. (a) Collaborative Tokenizer module converts item-level logits from CF models into token-level distributions. (b) Soft Label Alignment module fuses these CF-informed token probabilities with one-hot supervision to construct soft labels. Finally, the LLM is fine-tuned with a soft next-token prediction loss which incorporates collaborative signals.
  • Figure 2: Relation between collaborative consistency and performance. The x-axis refers to the hyperparameter $\alpha$. The dashed lines denote the collaborative consistency of three models.