Multi-LLM Token Filtering and Routing for Sequential Recommendation

Abstract: LLMs have recently shown promise in recommendation by providing rich semantic knowledge. While most existing approaches rely on external textual corpora to align LLMs with recommender systems, we revisit a more fundamental yet underexplored question: Can recommendation benefit from LLM token embeddings alone without textual input? Through a systematic empirical study, we show that directly injecting token embeddings from a single LLM into sequential recommenders leads to unstable or limited gains, due to semantic misalignment, insufficient task adaptation, and the restricted coverage of individual LLMs. To address these challenges, we propose MLTFR, a Multi-LLM Token Filtering and Routing framework for corpus-free sequential recommendation. MLTFR follows an interaction-guided LLM knowledge integration paradigm, where task-relevant token embeddings are selected via user-guided token filtering to suppress noisy and irrelevant vocabulary signals. To overcome the limitations of single-LLM representations, MLTFR integrates multiple LLM token spaces through a Mixture-of-Experts architecture, with a Fisher-weighted semantic consensus expert to balance heterogeneous experts and prevent domination during training. By jointly filtering informative tokens and aggregating complementary semantic knowledge across multiple LLMs, MLTFR enables stable and effective utilization of LLM token embeddings without textual inputs or backbone modification. Extensive experiments demonstrate that MLTFR consistently outperforms state-of-the-art sequential recommendation baselines and existing alignment methods. Our code is available at: https://github.com/ccwwhhh/MLTFR.

• The paper introduces MLTFR, a framework that selectively filters and routes LLM token embeddings to boost recommender performance.
• It employs a differentiable Gumbel-softmax token filtering and MoE-based aggregation to align intrinsic LLM semantics with task-specific needs.
• Empirical results across multiple datasets show significant improvements in HR@20 and NDCG@20 compared to traditional prompt-based LLM approaches.

Multi-LLM Token Filtering and Routing for Sequential Recommendation

Motivation and Problem Formulation

The integration of LLMs within recommender systems has primarily focused on exploiting external corpora to align semantic representations, typically relying on item or user textual information. However, this paradigm is fundamentally limited when high-quality textual corpora are unavailable, incomplete, or incur prohibitive maintenance overheads. The examined work challenges this dependency by proposing a corpus-free paradigm: direct utilization of the intrinsic token embedding spaces of pretrained LLMs as semantic augmentation resources for sequential recommenders. A systematic empirical analysis reveals that naive injection of such embeddings is destabilizing and only marginally beneficial, mainly due to large, noisy LLM vocabularies and semantic misalignment between token and task-specific domains.

The central hypothesis posited is that only a subset of LLM token embeddings, judiciously selected and aligned with user histories, can serve as efficient and robust semantic injectors in recommendation models, providing significant gains without recourse to textual features.

Figure 1: Task-relevant LLM token filtering effectively denoises semantic injection; empirical results show nontrivial gains for filtered LLM token embeddings compared to both random and unfiltered LLM embeddings on standard sequential recommendation backbones.

MLTFR Architecture

To realize this paradigm, the paper introduces MLTFR (Multi-LLM Token Filtering and Routing)—a modular plug-and-play MoE system that routes distilled token embeddings from multiple LLMs into sequential recommendation backbones. The principal innovations comprise two synergistic mechanisms: (1) an interaction-guided, differentiable token filtering strategy that identifies user-relevant tokens from each LLM vocabulary, and (2) an MoE integration pipeline leveraging both expert diversity and a Fisher-weighted semantic consensus expert to stabilize multi-LLM knowledge fusion.

The framework is architected as follows:

• Task-Relevant Token Filtering: For each LLM expert, token selection uses the user's sequence representation as a proxy for current interests. A temperature-controlled Gumbel-softmax router provides a differentiable approximation to hard Top-$K$ selection, operating in a projected alignment space between LLM embedding and recommender embedding domains. Only the selected tokens (of size $K$) are cross-attended with the sequence embedding.
• Semantic Integration Layer: Multi-head cross-attention fuses the selected LLM token embeddings with the backbone sequence representation, mitigating representational anisotropy and preserving compatibility with the underlying recommender.
• Gated Expert Aggregation: A parallel-gated MoE aggregates semantic matrices from $n$ LLM experts with item-level, learnable soft routing. This permits expert-specific specialization and sparse information flow.
• Semantic Consensus Expert: To suppress domination and conflicts from heterogeneous experts, Fisher Information is estimated for each expert; the shared expert is a frozen, Fisher-weighted convex combination of all expert parameters, ensuring consensus and regularization during training.

  Figure 2: MLTFR framework with per-expert LLM token filtering, parallel MoE expert routing, and Fisher-based semantic consensus expert fusion pipeline.

Experimental Analysis

A thorough empirical campaign on multiple public recommendation datasets (Amazon Office, Pantry, Music, RateBeer) and across strong backbone models (SASRec, BERT4Rec, FMLPRec, LinRec, CL4SRec, BSARec) demonstrates:

• Consistent performance improvements: MLTFR yields improved Hit Ratio (HR@20) and NDCG@20 in 21/24 model-dataset settings. The magnitude of improvement is maximized in BERT4Rec backbones (mean relative gain: 19.82%), indicating the synergy between bidirectional context modeling and LLM-based global semantics.
• Domain effect: Gains are domain-dependent, highest in Pantry (mean 12.69%) and lowest in Beer, reflecting the domain affinity of candidate LLMs.
• MoE scaling: Increasing the number of diverse LLM experts provides gains, though diminishing returns manifest with less relevant experts, underscoring the importance of expert coverage and alignment.

  Figure 3: MLTFR exhibits positive scaling trends with increased expert count on the Office dataset across six backbone models.

• Ablation of Semantic Consensus: Removal of the Fisher-based consensus expert degrades stability and overall performance, especially in heterogeneous or semantically distant expert sets. The effect is most pronounced in models with higher semantic capacity.

Black-box LLM Comparison

A direct comparison with prompt-based, text-conditioned black-box LLM recommenders (ChatGLM/glm4-plus) reveals a strong contrast: MLTFR-equipped ID-based models, operating without any text, achieve much higher top-1 accuracy (0.246 vs 0.121 for Office), even under fair candidate constraints. This demonstrates that, in absence of explicit item corpora, task-guided token selection from LLM embeddings is significantly more effective than direct LLM inference with text templates.

Theoretical and Practical Implications

The corpus-free, lightweight nature of MLTFR is particularly appealing for industrial recommendations where real-world item corpora are inconsistent or unavailable. The Fisher-based consensus mechanism prevents over-reliance on any one expert, making the solution robust under LLM heterogeneity and domain shift. The modular design allows for seamless insertion into existing recommenders, incurs only a modest computational overhead compared to full text-encoding approaches, and is amenable to parallelization.

A key theoretical insight is that LLM token embedding spaces can serve as generic semantic reservoirs, but only if filtered and routed via user-centric and model-centric mechanisms. MoE-based multi-LLM routing provides an effective medium for fusing non-overlapping semantic perspectives with minimal representational conflict.

The results contradict the implicit assumption in prior art that textual corpora are essential for semantic augmentation; this framework produces robust improvement without text and minimal backbone modifications.

Future Directions

Several lines of inquiry arise from these findings:

• Automated LLM expert selection and dynamic expert routing conditioned on user or domain properties.
• Extension beyond token embeddings to intermediate or higher-level LLM representations.
• Generalization to multi-modal and session-based recommendation, leveraging user- or context-specific expert assignment.
• Adaptive Fisher-weighted consensus with online model selection and regularization.

Conclusion

The proposed MLTFR framework demonstrates that corpus-free semantic augmentation for sequential recommendation is feasible and effective through multi-LLM token selection and gated expert routing. The design delivers consistent performance gains, strong robustness to expert heterogeneity, and outperforms black-box, prompt-based LLM recommenders under textless constraints. This constitutes a substantial advancement in the practical deployment of LLM competencies for industrial recommendation pipelines without incurring the cost or limitations of textual corpora.

(2604.18200)