Token-Controlled Re-ranking for Sequential Recommendation via LLMs

Wenxi Dai, Wujiang Xu, Pinhuan Wang, Dimitris N. MetaxasarXiv 2025 · arXiv · OpenReview

Abstract

The widespread adoption of Large Language Models (LLMs) as re-rankers is shifting recommender systems towards a user-centric paradigm. However, a significant gap remains: current re-rankers often lack mechanisms for fine-grained user control. They struggle to balance inherent user preferences with multiple attribute-based constraints, often resorting to simplistic hard filtering that can excessively narrow the recommendation pool and yield suboptimal results. To bridge this gap, we propose COREC, a novel token-augmented re-ranking framework that incorporates specific user requirements in co-creating the recommendation outcome. COREC empowers users to steer re-ranking results with precise and flexible control via explicit, attribute-based signals. The framework learns to balance these commands against latent preferences, yielding rankings that adhere to user instructions without sacrificing personalization. Experiments show that COREC: (1) exceeds state-of-the-art baselines on standard recommendation effectiveness and (2) demonstrates superior adherence to specific attribute requirements, proving that COREC enables fine-grained and predictable manipulation of the rankings.

Introduction

Most modern recommender systems operate on a "passive recipient" model, where algorithms analyze historical behavior to predict future interests. While LLMs have improved these predictions, they still struggle with direct user agency. When a user has a specific intent—such as finding electronics under $50—traditional systems often apply "hard filters." This binary approach can aggressively clear the candidate pool, potentially removing highly relevant items that only slightly miss a constraint. COREC addresses this by providing fine-grained control, balancing explicit user commands with latent historical preferences without breaking the recommendation engine.

Methodology

COREC (Token-Controlled Re-ranking) is a two-stage framework. First, a retrieval stage (e.g., SASRec) identifies a broad set of candidate items. Second, a re-ranking stage utilizes a fine-tuned Llama3-8B model via Low-Rank Adaptation (LoRA).

The COREC architecture: from input construction with control tokens to training with pairwise RankNet loss.

The key innovation is the introduction of Control Tokens. By expanding the model's vocabulary with special tokens like <price>0-10 or <brand>Disney, the framework treats user constraints as learnable embeddings rather than simple text prompts. The model is trained using a pairwise RankNet loss, teaching it to rank items based on a "control score" that accounts for both ground-truth relevance and adherence to explicit user constraints.

Key Results

COREC was evaluated on the Amazon 2018 dataset across categories such as Home and Kitchen, Electronics, and All Beauty. The framework demonstrated significant improvements over state-of-the-art baselines like LLaRA and TALLRec:

  • Recommendation Quality: COREC achieved a 17.3% to 40.4% improvement in standard metrics like NDCG (Normalized Discounted Cumulative Gain).
  • Controllability: Using a new metric, Control Precision (CP@K), COREC showed a 33.4% average improvement, indicating it is far more likely to show items matching requested attributes.
Performance comparison showing the significant gains in metrics like NDCG and Control Precision (CP) after fine-tuning with COREC.

Unlike traditional filtering, which becomes too restrictive as constraints increase, COREC maintains high recommendation quality by finding the best compromise among multiple overlapping requirements.

Conclusion

COREC transforms LLM re-rankers from passive observers into active collaborators. By treating constraints as tokens rather than binary filters, the model learns a nuanced reasoning path that respects both user intent and historical preference. Future work will explore scaling this to more complex, hierarchical attributes and optimizing the computational overhead for real-time deployment.

Summary updated

Abstract

The widespread adoption of Large Language Models (LLMs) as re-rankers is shifting recommender systems towards a user-centric paradigm. However, a significant gap remains: current re-rankers often lack mechanisms for fine-grained user control. They struggle to balance inherent user preferences with multiple attribute-based constraints, often resorting to simplistic hard filtering that can excessively narrow the recommendation pool and yield suboptimal results. This limitation leaves users as passive recipients rather than active collaborators in the recommendation process. To bridge this gap, we propose COREC, a novel token-augmented re-ranking framework that incorporates specific user requirements in co-creating the recommendation outcome. COREC empowers users to steer re-ranking results with precise and flexible control via explicit, attribute-based signals. The framework learns to balance these commands against latent preferences, yielding rankings that adhere to user instructions without sacrificing personalization. Experiments show that COREC: (1) exceeds state-of-the-art baselines on standard recommendation effectiveness and (2) demonstrates superior adherence to specific attribute requirements, proving that COREC enables fine-grained and predictable manipulation of the rankings.
BibTeX
@misc{dai2025token,
  author = {Wenxi Dai and Wujiang Xu and Pinhuan Wang and Dimitris N. Metaxas},
  title = {{Token-Controlled Re-ranking for Sequential Recommendation via LLMs}},
  year = {2025},
  eprint = {2511.17913},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.17913},
}
Figure