2022
Best Paper
Neural Corpus Indexer
Generative retrieval that indexes a corpus end-to-end inside model weights.
Core contributor · NeurIPS
Haonan Wang
王淏楠Principal Research Scientist Tencent Hunyuan@Singapore
I am a Principal Research Scientist at Tencent Hunyuan, where I work on reinforcement learning (RL) for multimodal generation (video, image, text, etc.) and lead UniRL, our open-source RL framework for post-training foundation models toward agentic multimodal generation.
Concurrently, I am completing my PhD at the National University of Singapore, advised by Prof. Kenji Kawaguchi. Previously a PhD student at UIUC with Prof. Jingrui He, and B.S. in CS & Statistics at UIUC mentored by Prof. Jiawei Han.
My research builds systems that reason, plan, and learn from interaction — spanning long-context training, world models, LLM reasoning, autonomous agents, and reinforcement learning for unified multimodal generation.
Feel free to reach out if you'd like to chat about research or collaborate.
Research Arc
2024
AnchorAttention
Diagnosed why BFloat16 silently breaks RoPE under long context — and patched it.
First author · TMLR
2025
Next-Event Prediction
A future-prediction objective for video — world models that anticipate events, not just tokens.
Co-first author · ICLR
2025
Reasoning Systems
Parallel-decoding aggregation and demo-driven supervision for stronger reasoners.
First author · ICLR
2025
MiroThinker
Open-source deep-research agent scaling model, context, and interaction.
Open source · MiroMind
Now
UniRL
Open-source RL framework for post-training foundation models toward agentic multimodal generation.
Project lead · Tencent Hunyuan
