Title: GFlowGR: Adapting Generative Recommendation Models via Generative Flow Networks

Abstract: Generative Recommendation (GR) systems, typically combining item tokenizers with generative Large Language Models (LLMs), have achieved significant success across diverse applications. While current research predominantly focuses on enhancing item tokenizers or refining LLM decoding mechanisms to boost performance, the crucial fine-tuning phase—necessary for aligning LLMs with recommendation data—has received limited attention. Existing methods mostly depend on either the next-token prediction loss used in Supervised Fine-Tuning (SFT) or recommendation-specific Direct Preference Optimization (DPO). However, both strategies overlook the exploration of potential positive but unobserved samples, a issue known as the exposure bias problem. To address this limitation, this study models GR as a multi-step generation process and introduces GFlowGR, a fine-tuning framework grounded in GFlowNets. This framework incorporates collaborative knowledge from conventional recommender systems to develop an adaptive trajectory sampler and a robust reward model. By utilizing the inherent diverse generation capabilities of GFlowNets, supplemented by sampling and heuristic weighting techniques, GFlowGR offers a promising solution to reduce exposure bias. Comprehensive experiments on two real-world datasets using two distinct GR backbones demonstrate the effectiveness and robustness of the proposed GFlowGR approach.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC