Sculpting User Preferences for Recommendation with Positive-Negative Diffusion Guidance

ICLR 2026 Conference Submission10573 Authors

Keywords: Diffusion Models, Recommender Systems, Sequential Recommendation, Negative Guidance

TL;DR: We propose SteerRec, a diffusion model for recommendation that uses a novel positive-negative guidance mechanism, enabled by an alignment-based triplet loss, to achieve more precise user preference generation.

Abstract: Diffusion models are emerging as a powerful generative paradigm for sequential recommendation, demonstrating a remarkable ability to model complex user-item interaction dynamics. Despite their strong modeling ability, most diffusion-based recommenders face limited generative control because the standard classifier-free guidance derives its repulsive signal from a global and user-agnostic unconditional prior, which prevents the model from directly exploiting negative feedback at inference. A natural solution is to replace the unconditional prior with user-aware negative conditions. However, this is challenging because, unlike in text-to-image tasks where negative prompts acquire stable semantics from a pre-trained text encoder, item embeddings in recommendation are learned dynamically. As a result, a ''negative condition'' is not guaranteed to provide effective repulsive guidance unless the model is explicitly trained to recognize it as a signal for avoidance. To enable effective and steerable negative guidance in diffusion recommenders, we propose **SteerRec**, a novel framework built upon two core innovations. At inference, we introduce Positive-Negative Guidance (PNG) inference mechanism, which replaces the generic unconditional prior with a user-aware negative condition. To ensure the negative condition provides meaningful repulsive guidance in the dynamic embedding space, we design a Guidance Alignment Triplet Loss (GAL). The GAL is a margin-based objective that explicitly aligns the training process with PNG by ensuring the model’s prediction under a positive condition is closer to the target item than its prediction under a negative condition. Extensive experiments on three widely used public benchmarks provide strong empirical evidence for the effectiveness of SteerRec. Our implementation is available at \url{https://anonymous.4open.science/r/SteerRec-5D70}.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 10573