Peng Yan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Peng Yan has a total of 27 co-authors. Here are the top co-authors by number of shared publications:
- Biao Tang
Meituan, Shanghai, China
1 shared publication - Chong Peng
Meituan, Shanghai, China
1 shared publication - Disen Hu
School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China
1 shared publication - Fuli Feng
National University of Singapore, Singapore, Singapore
1 shared publication - Hao Yang
Meituan, Shanghai, China
1 shared publication - Hao Yin
Mei Tuan, Beijing, China
1 shared publication
Latest Publications
- Open Access
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 6207-6218
Learning from Unreachable Rewards: Hint-Conditioned Reinforcement Learning for Generative Recommendation
Semantic-ID generative recommenders represent each item as a short sequence of discrete semantic tokens and predict the next item by autoregressively generating this token sequence. This paradigm enables a unified generation interface for item IDs, histories, and item text, but it also creates a structured optimization bottleneck during reward-based post-training: when an early semantic token enters the wrong branch of the item-token space, finite rollout groups rarely reach the ground-truth item, so group-relative optimization receives identical zero rewards and produces no useful advantage. We propose Hint-Conditioned Generative Recommendation (\method), a semantic-ID generative recommendation framework that recovers learning signal for such hard training instances. HCGRec diagnoses each instance with checkpoint rollouts and supplies a minimal target-prefix hint only when the current generator cannot reach the correct item. The model then generates the unhinted suffix under the hinted semantic branch, turning zero-reward groups into informative comparisons over item-token completions. Hinting also changes token identity: hinted prefix tokens are oracle-provided item context, while unhinted suffix tokens are sampled generation actions. We therefore introduce hint-aware credit decomposition, using supervised learning to preserve item-semantic and prefix-structure alignment for hinted tokens and GRPO to optimize the sampled suffix. Experiments on sequential recommendation benchmarks show that HCGRec substantially improves over supervised fine-tuning and vanilla reward-based post-training, while reducing zero-advantage training samples from over 70% to below 20%. The code is accessible at https://github.com/WncFht/GRec.
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 6207-6218
Open AccessLearning from Unreachable Rewards: Hint-Conditioned Reinforcement Learning for Generative Recommendation
Semantic-ID generative recommenders represent each item as a short sequence of discrete semantic tokens and predict the next item by autoregressively generating this token sequence. This paradigm enables a unified generation interface for item IDs, histories, and item text, but it also creates a structured optimization bottleneck during reward-based post-training: when an early semantic token enters the wrong branch of the item-token space, finite rollout groups rarely reach the ground-truth item, so group-relative optimization receives identical zero rewards and produces no useful advantage. We propose Hint-Conditioned Generative Recommendation (\method), a semantic-ID generative recommendation framework that recovers learning signal for such hard training instances. HCGRec diagnoses each instance with checkpoint rollouts and supplies a minimal target-prefix hint only when the current generator cannot reach the correct item. The model then generates the unhinted suffix under the hinted semantic branch, turning zero-reward groups into informative comparisons over item-token completions. Hinting also changes token identity: hinted prefix tokens are oracle-provided item context, while unhinted suffix tokens are sampled generation actions. We therefore introduce hint-aware credit decomposition, using supervised learning to preserve item-semantic and prefix-structure alignment for hinted tokens and GRPO to optimize the sampled suffix. Experiments on sequential recommendation benchmarks show that HCGRec substantially improves over supervised fine-tuning and vanilla reward-based post-training, while reducing zero-advantage training samples from over 70% to below 20%. The code is accessible at https://github.com/WncFht/GRec.
- Open Access
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 1337-1345
Geometric Gradient Divergence Modulation for Imbalanced Multimodal Learning
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 1337-1345
Open AccessGeometric Gradient Divergence Modulation for Imbalanced Multimodal Learning
- Open Access
- 22 April 2025
- Published by Association for Computing Machinery (ACM)
- p. 264-274
Personalized Image Generation with Large Multimodal Models
- 22 April 2025
- Published by Association for Computing Machinery (ACM)
- p. 264-274
Open AccessPersonalized Image Generation with Large Multimodal Models
- Open Access
- 14 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 53-57
Large Scale Hierarchical User Interest Modeling for Click-through Rate Prediction
With the explosive growth of online information, recommender systems have emerged as indispensable tools for navigating the complexities of content generation, discovery and consumption. The RecSys 2024 Challenge, organized by Ekstra Bladet, provides a comprehensive dataset and a robust news recommendation evaluation framework for tackling multifaceted challenges including modeling user preferences based on implicit behavior, accounting for the influence of the news agenda on user interests, and managing the rapid decay of news items. In this paper, we propose a novel Hierarchical User Interest Modeling (HUIM) approach to this challenge, leveraging both long-term invariant interests and short-term rapidly changing interests. Specifically, we utilize the multimodal representations along with the side information of items to distill the long-standing interests from user historical behaviors. By analyzing real-time context and user behavioral path patterns, we identify their fine-grained instant interests. Finally, an interest fusion network is proposed to adaptively fuse the long-term and short-term interests by contrasting the query-aware fine-grained behaviors with query-level cross entropy loss. Our team, BlackPearl, achieved a score of 0.8815 and ranked 2nd place on the final leaderboard.
- 14 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 53-57
Open AccessLarge Scale Hierarchical User Interest Modeling for Click-through Rate Prediction
With the explosive growth of online information, recommender systems have emerged as indispensable tools for navigating the complexities of content generation, discovery and consumption. The RecSys 2024 Challenge, organized by Ekstra Bladet, provides a comprehensive dataset and a robust news recommendation evaluation framework for tackling multifaceted challenges including modeling user preferences based on implicit behavior, accounting for the influence of the news agenda on user interests, and managing the rapid decay of news items. In this paper, we propose a novel Hierarchical User Interest Modeling (HUIM) approach to this challenge, leveraging both long-term invariant interests and short-term rapidly changing interests. Specifically, we utilize the multimodal representations along with the side information of items to distill the long-standing interests from user historical behaviors. By analyzing real-time context and user behavioral path patterns, we identify their fine-grained instant interests. Finally, an interest fusion network is proposed to adaptively fuse the long-term and short-term interests by contrasting the query-aware fine-grained behaviors with query-level cross entropy loss. Our team, BlackPearl, achieved a score of 0.8815 and ranked 2nd place on the final leaderboard.