Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ThinkRec: Thinking-based recommendation via LLM

Qihang Yu, Kairui Fu, Zheqi Lv, Shengyu Zhang, Xinhui Wu, Chen Lin, Feng Wei, Bo Zheng, Fei Wu

TL;DR

ThinkRec presents a thinking-based framework that pushes LLM-based recommender systems toward explicit, verifiable reasoning by augmenting item metadata, synthesizing reasoning traces, and training with reasoning data. It introduces reasoning-augmented training and a dynamic, instance-wise LoRA expert fusion to adapt reasoning paths to individual users, improving both accuracy and interpretability. Empirical results on three real-world datasets show significant improvements in AUC and in the quality of generated explanations, along with strong cross-domain generalization and favorable inference efficiency. The work advances recommender systems by bridging collaboration signals with language-model reasoning, enabling more faithful, personalized, and transparent recommendations.

Abstract

Recent advances in large language models (LLMs) have enabled more semantic-aware recommendations through natural language generation. Existing LLM for recommendation (LLM4Rec) methods mostly operate in a System 1-like manner, relying on superficial features to match similar items based on click history, rather than reasoning through deeper behavioral logic. This often leads to superficial and erroneous recommendations. Motivated by this, we propose ThinkRec, a thinking-based framework that shifts LLM4Rec from System 1 to System 2 (rational system). Technically, ThinkRec introduces a thinking activation mechanism that augments item metadata with keyword summarization and injects synthetic reasoning traces, guiding the model to form interpretable reasoning chains that consist of analyzing interaction histories, identifying user preferences, and making decisions based on target items. On top of this, we propose an instance-wise expert fusion mechanism to reduce the reasoning difficulty. By dynamically assigning weights to expert models based on users' latent features, ThinkRec adapts its reasoning path to individual users, thereby enhancing precision and personalization. Extensive experiments on real-world datasets demonstrate that ThinkRec significantly improves the accuracy and interpretability of recommendations. Our implementations are available at https://github.com/Yu-Qi-hang/ThinkRec.

ThinkRec: Thinking-based recommendation via LLM

TL;DR

ThinkRec presents a thinking-based framework that pushes LLM-based recommender systems toward explicit, verifiable reasoning by augmenting item metadata, synthesizing reasoning traces, and training with reasoning data. It introduces reasoning-augmented training and a dynamic, instance-wise LoRA expert fusion to adapt reasoning paths to individual users, improving both accuracy and interpretability. Empirical results on three real-world datasets show significant improvements in AUC and in the quality of generated explanations, along with strong cross-domain generalization and favorable inference efficiency. The work advances recommender systems by bridging collaboration signals with language-model reasoning, enabling more faithful, personalized, and transparent recommendations.

Abstract

Recent advances in large language models (LLMs) have enabled more semantic-aware recommendations through natural language generation. Existing LLM for recommendation (LLM4Rec) methods mostly operate in a System 1-like manner, relying on superficial features to match similar items based on click history, rather than reasoning through deeper behavioral logic. This often leads to superficial and erroneous recommendations. Motivated by this, we propose ThinkRec, a thinking-based framework that shifts LLM4Rec from System 1 to System 2 (rational system). Technically, ThinkRec introduces a thinking activation mechanism that augments item metadata with keyword summarization and injects synthetic reasoning traces, guiding the model to form interpretable reasoning chains that consist of analyzing interaction histories, identifying user preferences, and making decisions based on target items. On top of this, we propose an instance-wise expert fusion mechanism to reduce the reasoning difficulty. By dynamically assigning weights to expert models based on users' latent features, ThinkRec adapts its reasoning path to individual users, thereby enhancing precision and personalization. Extensive experiments on real-world datasets demonstrate that ThinkRec significantly improves the accuracy and interpretability of recommendations. Our implementations are available at https://github.com/Yu-Qi-hang/ThinkRec.

Paper Structure

This paper contains 39 sections, 10 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. LLM-based Recommendation
  4. Method
  5. Preliminary
  6. Thinking Enhanced Recommendation
  7. Reasoning data construction
  8. Thinking activation
  9. Recommendation Experts Fusion
  10. Base Expert fine-tuning
  11. Instance Wise Expert Fusion
  12. Staged Training
  13. Experiments
  14. Experimental Setup
  15. Datasets.
  16. ...and 24 more sections

Figures (4)

  • Figure 1: (a) shows the composition of user interaction data. (b) and (c) illustrate previous LLM-based recommendations and our ThinkRec, respectively. (d) compares ThinkRec with baselines in three real-world datasets.
  • Figure 2: Overview of the proposed ThinkRec. Keywords are extracted from the description of items with the pretrained PolyLM-Qwen-7B. Reasoning data is synthesized with a reasoning model from a sample of training data. Global LoRA is trained with total data, and base LoRAs are tuned on it with data grouping by user feature. The used LoRA, fusioned or single, is determined by the similarity between the user feature and LoRA representations.
  • Figure 3: The influence of performance with the number of experts on Book (left panel) and Yelp (right panel).
  • Figure 4: The influence of performance on the accuracy of grouping (Cohen's d of grouped datasets).