CaLR: Context-Aware Latent Reasoning for Sequential Recommendation

Abstract

Sequential recommendation aims to model users' evolving preferences from their historical interactions. Recent latent reasoning methods improve sequential recommenders by extending inferencetime computation in continuous representation space, refining user representations without generating explicit reasoning text. However, existing methods mainly reason over the encoded history itself and make limited use of contextual feedback from users' recent interactions. Such feedback, especially positive and negative signals in the latest session, can not only complement the long-term user representation but also provide directional guidance for latent reasoning by indicating which preference patterns should be strengthened or suppressed. In this paper, we propose Context-Aware Latent Reasoning (CaLR), a model-agnostic framework that grounds latent reasoning in session-level context. From the input perspective, CaLR selects informative positive and negative items from users' recent sessions as reasoning anchors. From the architecture perspective, it injects these anchors into each reasoning step through a Cross-Attention-based module. From the optimization perspective, it introduces Process Supervision Loss (PSL) to supervise intermediate reasoning states according to anchor feedback polarity. Experiments on two real-world public benchmarks show that CaLR consistently improves multiple sequential recommendation backbones and outperforms strong latent reasoning baselines. Ablation and further analyses verify the effectiveness of each component in CaLR. Moreover, a seven-day online A/B test on Kuaishou Lite demonstrates its effectiveness in real-world industrial deployment. Our code is available at https://github.com/Ak4ne-windsky/CaLR.