LACEREC: Controllable Sequential Recommendation with Weak-Signal Enhancement

Sequential recommendation has achieved remarkable success in modeling user dynamics, yet controllability remains a fundamental challenge, particularly for weak signals such as niche or high-quality items. To address this issue, we propose Layerwise Attentive Control Enhancement (LACERec), a controllable sequential recommendation framework. LACERec consists of two key components: (i) layerwise cross-attention injection, where control tokens are introduced at every Transformer layer to guide representation learning; and (ii) dual-head projection with dynamic mixing, which provides a dedicated path to enhance weak-signal categories while preserving overall ranking performance. Experiments on three real-world datasets demonstrate that LACERec significantly improves weak-signal metrics while maintaining competitive overall accuracy, thereby alleviating popularity bias and establishing a robust paradigm for controllable sequential recommendation.