New method improves generative recommendation by learning from user history

Rethinking Latent Reasoning for Generative Recommendation: Beyond Teacher-Generated CoT and Recurrent Hidden-State Feedback

Information Retrieval

Summary

Generating recommendations often depends on teaching AI models to think through steps, but creating these teaching examples is costly and slow. The authors found that analyzing users’ past behavior was the most helpful part in making accurate predictions. They created a new approach called PHiT that learns from these past behaviors directly, avoiding complicated feedback loops. This method is faster to train and performs better on multiple recommendation tests.

What this means in practice

  • For ecommerce platform engineers: Improve product recommendations by integrating user history cues without costly supervision steps.$Commercial implications: This paper enables more accurate and efficient recommendation engines which can be offered as a commercial product improvement for ecommerce sites.
  • For streaming service developers: Enhance content suggestion systems by using prefilled history-informed tokens for faster user preference prediction.

Authors

Yiwen Chen, Fuwei Zhang, Zehao Chen, Deqing Wang, Hehan Li, Peizhi Xu, Hanmeng Liu, Shuanglong Li, Xin Pei, Fuzhen Zhuang, Zhao Zhang

Abstract

Latent reasoning in generative recommendation faces two practical barriers: constructing teacher-generated chain-of-thought (CoT) supervision is costly, and recurrent hidden-state feedback requires a specialized execution loop. We revisit both the supervision source and the computation interface. A step-level diagnosis measures how recommendation CoT supports next-item prediction and trace coherence. In the analyzed traces, history analysis is the only step with positive mean contributions on both dimensions, while later preference and intent steps have negative mean answer contributions. Guided by this finding, we propose PHiT (Prefilled, History-Informed Thought Tokens), which learns intermediate states from semantic targets extracted directly from observed item histories. PHiT aggregates cached sparse item encodings to construct these targets and supervises the contextual states of distinct vocabulary thought tokens. All thought tokens are supplied together as ordinary input IDs during a single causal prefill, removing the need for recurrent final-state feedback. With supervised fine-tuning alone, PHiT outperforms the compared baselines across four recommendation benchmarks. Controlled ablations support the value of input-matched semantic supervision, while cost measurements show fast target construction with training time comparable to the recommendation-only thought-token baseline.