About Me 👨‍🔬

I am currently a Assistant Professor at the Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), affiliated with Shenzhen University.

My current research interests🔬 focus on RWKV, a efficient linear attention architecture for language modeling.

My work has been published at leading AI venues, including EMNLP, COLING, COLM, and AAAI, with 20+ papers and 1,700+ citations.

Work Experience

• 2023.08 - Now, Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Assistant Professor
• 2017.08 - 2022.11, Tencent Machine Learning Platform (Now HunYuan Team), Senior Research Scientist

Educations 📖

• 2013.01 - 2017.05, Ph.D., National University of Singapore, Singapore.
• 2008.08 - 2012.07, B.Eng., Harbin Institute of Technology, Harbin.

News 🔥

• 2026.02: 🎉 SceneRAG is accepted by ICASSP 2026
• 2025.12: 🎉 VisualRWKV-HM is accepted by Information Fusion (IF=15.5)
• 2025.11: 🎉 PSPO is accepted by AAAI 26
• 2025.03: 🎉 RWKV-UI is accepted by ICME 2025
• 2024.12: 🎉 VisualRWKV is accepted by COLING 2025, check out the chinese introduction video on Bilibili 🎬

Selected Projects 🏭

2026 Large-Scale LLM Pre-Training, Mid-Training, and Multimodal Training with RWKV

• Served as a core technical leader in the development of next-generation large language model architectures in the RWKV series, including linear-attention models (RWKV-4/5/6/7), hybrid-attention models (RWKV-X), and vision-language models (VisualRWKV).
• Made major contributions across the full training pipeline, including large-scale pre-training, mid-training, and multimodal training, with primary ownership of the multimodal training direction.
• Led or co-led key efforts in scaling models beyond 10B parameters and training on approximately 8T tokens, covering architecture innovation, training methodology, and large-scale system optimization.
• Advanced a novel architecture that unifies Transformer-style parallelizable training with RNN-style efficient inference, making RWKV a strong candidate for future foundation model and AGI-oriented architectures.

2026 LLM Post-Training Optimization with PSPO (Accepted by AAAI-26)

• Led/participated in the design of an efficient reinforcement fine-tuning method for large language models, addressing training instability and poor data utilization in GRPO-based post-training.
• Introduced EMA-based reward smoothing and priority-based prompt sampling, improving the stability of policy updates and increasing the reuse value of high-impact prompts without storing full trajectories.
• Boosted training effectiveness over GRPO baselines with up to 5%+ absolute improvement in some settings, while also improving convergence speed.
• Open-sourced the project and validated it on Qwen2.5-Math-1.5B/7B and other model variants across multiple mathematical reasoning benchmarks, including AIME24, MATH500, AMC23, Minerva Math, and OlympiadBench.

Selected Publications 📝

AAAI 2026

PSPO: Prompt-Level Prioritization and Experience-Weighted Smoothing for Efficient Policy Optimization
Xinxin Zhu, Haowen Hou, Ruichong Zhang, Nianbo Zeng, Yulin Peng, Jiongfeng Fang, Fei Yu, Ying Tiffany He

Code | Paper

Information Fusion

VisualRWKV-HM: Enhancing linear visual-language models via hybrid mixing
Haowen Hou, Fei Ma, Zihang Li, Fei Richard Yu

Code | Paper

COLING 2025

VisualRWKV: Exploring Recurrent Neural Networks for Visual Language Models
Haowen Hou and Peigen Zeng and Fei Ma and Fei Richard Yu

Code | Paper | Video | BibTeX

COLM 2025

RWKV-7 “Goose” with Expressive Dynamic State Evolution
Bo Peng, Ruichong Zhang, Daniel Goldstein, Eric Alcaide, Haowen Hou，et al.

Paper | HuggingFace | Code

ICME 2025

RWKV-UI: UI Understanding with Enhanced Perception and Reasoning
Jiaxi Yang, Haowen Hou

Paper

IEEE SPL

MLLM-TA: Leveraging Multimodal Large Language Models for Precise Temporal Video Grounding
Yi Liu, Haowen Hou(co-first author), Fei Ma, Shiguang Ni, Fei Richard Yu

Paper

EAAI

BagFormer: Better cross-modal retrieval via bag-wise interaction
Haowen Hou, Xiaopeng Yan, Yigeng Zhang

Paper

COLM 2024

Eagle and Finch: RWKV with Matrix-Valued States and Dynamic Recurrence
Bo Peng, Eric Alcaide, Quentin Anthony, …, Haowen Hou, et al.

Paper | HuggingFace | Code

EMNLP 2023

RWKV: Reinventing RNNs for the Transformer Era
Bo Peng, Eric Alcaide, Quentin Anthony, …, Haowen Hou, et al.

Paper | HuggingFace | Code

Other Publications 📚

• arXiv EmbeddingRWKV: State-Centric Retrieval with Reusable States
• arXiv RWKV-X: A Linear Complexity Hybrid Language Model
• arXiv RWKV-TS: Beyond Traditional Recurrent Neural Network for Time Series Tasks

Honors and Awards 🎖

• 2018, Shenzhen High-Level Overseas Talent
• 2013, NUS SMA3 Scholarship
• 2012, NUS NGS Scholarship
• 2007, First Prize, National Olympiad in Chemistry in Provinces (Top 1%)

Invited Talks 💬

• 2025.03, RWKV: Next-Gen Model Architecture, NVIDIA GTC 2025
• 2024.12, RWKV: Next-Gen Model Architecture, Future Medicine Conference
• 2024.10, RWKV: Next-Gen Model Architecture, China National Conference on Social Media Processing 2024

Teaching 🧑‍🏫

• 2024 Fall Advanced Algorithms

Academic Service 🎓

• 2024.06 - Now, Journal Reviewer, IEEE Signal Processing Letters
• 2025.01 - Now, Conference Reviewer, IEEE International Conference on Multimedia & Expo (ICME)