Verifiable Reasoning for LLM-based Generative Recommendation

TL;DR

This work proposes a novel reasoning with verification paradigm, which interleaves reasoning with verification to provide reliable feedback, guiding the reasoning process toward more faithful user preference understanding, and proposes an effective implementation called VRec.

Abstract

Reasoning in Large Language Models (LLMs) has recently shown strong potential in enhancing generative recommendation through deep understanding of complex user preference. Existing approaches follow a {reason-then-recommend} paradigm, where LLMs perform step-by-step reasoning before item generation. However, this paradigm inevitably suffers from reasoning degradation (i.e., homogeneous or error-accumulated reasoning) due to the lack of intermediate verification, thus undermining the recommendation. To bridge this gap, we propose a novel \textbf{\textit{reason-verify-recommend}} paradigm, which interleaves reasoning with verification to provide reliable feedback, guiding the reasoning process toward more faithful user preference understanding. To enable effective verification, we establish two key principles for verifier design: 1) reliability ensures accurate evaluation of reasoning correctness and informative guidance generation; and 2) multi-dimensionality emphasizes comprehensive verification across multi-dimensional user preferences. Accordingly, we propose an effective implementation called VRec. It employs a mixture of verifiers to ensure multi-dimensionality, while leveraging a proxy prediction objective to pursue reliability. Experiments on four real-world datasets demonstrate that VRec substantially enhances recommendation effectiveness and scalability without compromising efficiency. The codes can be found at https://github.com/Linxyhaha/Verifiable-Rec.

Figures (11)

• Figure 1: Paradigm comparisons between (a) typical LLM-based recommendation, (b) reasoning for LLM-based recommendation, and (c) our proposed verifiable reasoning for LLM-based recommendation.
• Figure 2: Illustration of verification step in "reason-verify-recommend" paradigm. The multi-aspect verification ensures the verification diversity, while confidence-based adjustment emphasizes verification certainty.
• Figure 3: (a) illustrates the t-SNE visualization of the representations of the latent reasoning and the target item of LatentR$^3$ with 2 reasoning steps. (b) demonstrates continuous performance degradation as the reasoning step increases under the reason-then-recommend paradigm.
• Figure 4: Overview of VRec. (a) shows the design of the mixture of verifiers, including a personalized router, and a set of verifiers for different dimensions. We demonstrate two verifiers as an example for demonstration clarity. (b) shows the two-stage training strategy of VRec, including verifier pre-training and verifiable reasoning fine-tuning.
• Figure 5: Ablation study of VRec on CDs dataset.