Nov 2026· Proceedings of the 35th ACM International Conference on Information and Knowledge Management· pp. 1466-1477· 0 citations· 22 references
Abstract
Large Language Models (LLMs) have demonstrated significant potential in recommendation systems; however, employing Reinforcement Learning (RL) to align them with ranking metrics in the query-generation formulation of large-scale recommendation remains a critical challenge. Current RL-based approaches often suffer from reward sparsity and generating too general queries. In this paper, we propose TAG-Rec, a framework that grounds generative recommendation via Trigger-Anchored Group Relative Policy Optimization (GRPO). Built upon a verifiable-anchor training principle, TAG-Rec enforces a ''Trigger-First'' generation strategy in which the model explicitly identifies a salient historical item as a verifiable anchor before generating a search query, structurally transforming open-ended generation into grounded reasoning. TAG-Rec combines three components: (i) Trigger-Grounded Data Curation via reachability verification, which concentrates training supervision on samples that admit a verifiable historical anchor; (ii) the Trigger-First generation strategy that anchors generation in a salient prior interaction; and (iii) a Trigger-aware Reward that supplies dense intermediate feedback to mitigate reward sparsity. On real-world sequential recommendation benchmarks, TAG-Rec achieves up to 23.6% and 13.8% average improvement over strong SFT and RL baselines, with notable gains on long-tail and novelty metrics.
Generative recommendation promises to simplify conventional multi-stage recommender systems by generating item Semantic IDs (SIDs) autoregressively. However, two obstacles limit its effective- ness in industrial settings: interaction-only supervision underuses the semantic knowledge encoded in pretrained language model...
Zhuo-Hang Jiang, Yu-Xin Chen, Hao-Hao Qu et al.· 0 citations
Large language models (LLMs) have demonstrated strong reasoning capabilities, motivating their use as the backbone of foundation recommendation models (FRMs). Existing methods enhance recommendations through explicit Chain-of-Thought (CoT) reasoning under a Think-then-Answer paradigm. However, explicit CoT incurs subst...
Hao Jiang, Pei Du, Peng-Fei Yao et al.· 0 citations
Retrieval-Augmented Generation (RAG) improves Large Language Models (LLMs) by incorporating external documents, but its effectiveness degrades when retrieved passages are irrelevant, incomplete, or inconsistent with parametric knowledge. We present LPDG-RAG, a post-training framework that extends reasoning-based knowle...
Hai-Feng Liu, Rui-Qi Zhao, Nan Zhao et al.· IEEE Transactions on Audio,...· 0 citations
List Composer is introduced, a decoupled module that transforms pointwise candidate pools into deliverable exposure lists within generative recommender systems within generative recommender systems, and achieves consistent improvements in recommendation quality and serving efficiency.
Zhe-Li Zhou, Si-Thanh Dong, Wei-Wen Liu et al.· Proceedings of the 35th ACM...· 0 citations
Generative recommendation typically operates under the User-Induced Recommendation (UIR) setting, predicting target items by autoregressively generating semantic IDs as a passive continuation of long-term user history. However, this paradigm overlooks Trigger-Induced Recommendation (TIR), an essential industrial scenar...
Mu-Yu Zou, Hai-Bo Xing, Shun-Ye Tang et al.· 0 citations
LLM-based recommender systems increasingly generate chain-of-thought (CoT) text before each item prediction to activate the LLM's reasoning capacity, costing tens to hundreds of extra tokens per recommendation at inference. We ask whether inference-time reasoning artifacts (textual or latent) can be replaced by trainin...
Hong-Yang Su, Bei-Bei Kong, Lei Cheng et al.· Proceedings of the 35th ACM...· 1 citation
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.