Test1
2025-02-28
LLMsCoT DistillationModel Compression
How granularity, supervision format, and teacher models affect CoT distillation into smaller LMs across 7 benchmarks.
Read more
Hi, I’m Yanjun Chen.
PhD in RLHF & Embodied AI. | INTJ.
Builder of thinking agents.
Let’s explore minds that learn.
About Me
I'm a PhD student based in Hong Kong 🇭🇰, originally from China 🇨🇳. As an INTJ thinker and lifelong learner, I explore the frontiers of Artificial Intelligence with a special focus on RLHF (Reinforcement Learning with Human Feedback) and Embodied AI 🤖. My mind is always seeking structure, clarity, and elegant solutions.
📍 Location
Based in Hong Kong
🎯 MBTI
INTJ – The Architect
🎓 Education
PhD @ HK PolyU
💻 Programming
- Python 🐍
- C/C++ ⚙️
🌐 Languages
- Chinese 🇨🇳
- English 🇬🇧
- Japanese 🇯🇵
🎨 Interests
Table Tennis 🏓Video Games 🎮KTV 🎤Science & Tech 📖
Research
The Accuracy Paradox in RLHF: When Better Reward Models Don't Yield Better Language Models
We uncover a paradox where moderately accurate reward models outperform stronger ones in RLHF training. This challenges the assumption that better reward models yield better langua...
RLHFAlignmentReward Models
EMNLP 2024Cited 1×
Integrating Chain-of-Thought for Multimodal Alignment: A Study on 3D Vision-Language Learning
We explore integrating CoT into 3D vision-language alignment, showing significant gains through structured reasoning with our 3D-CoT benchmark.
MultimodalChain-of-Thought3D Reasoning
arXiv 2025Cited 0×
Corrected Soft Actor Critic for Continuous Control
We improve SAC by correcting action sampling bias introduced by tanh, achieving better convergence and performance on standard benchmarks.
Reinforcement LearningSACControl
arXiv 2024Cited 0×
Instruction-Tuned LLMs Succeed in Document-Level MT Without Fine-Tuning—But BLEU Turns a Blind Eye
We show instruction-tuned LLMs excel in docMT without fine-tuning. BLEU fails to capture improvements, and GPT-4 proves to be a better evaluator.
LLMsDocument MTEvaluation
arXiv 2024Cited 2×
Unveiling the Key Factors for Distilling Chain-of-Thought Reasoning
We dissect how granularity, supervision format, and teacher models affect CoT distillation into small language models across 7 datasets.
LLMsCoT DistillationModel Compression
arXiv 2025Cited 4×
Breaking the Pre-Planning Barrier: Real-Time Adaptive Coordination of Mission and Charging UAVs Using Graph RL
We introduce HGAM, a novel heterogeneous graph-based multi-agent RL model that enables real-time UAV coordination without pre-planned paths.
Multi-agent RLGraph NetworksUAVs
arXiv 2025Cited 0×
📝 Blog
View All →Test3
2024-03-12
RLHFPhilosophyAI Ethics
Reinforcement Learning from Human Feedback isn’t just a technical optimization, but a philosophical imperative in agent design.
Read more
Test2
2024-01-25
Embodied AINeuroscienceRL
Why grounding intelligence in a physical world is essential for the next generation of agents.
Read more
Contact
I'm always open to collaboration, interesting conversations, or just sharing ideas over coffee ☕. Whether it's about reinforcement learning, embodied AI, or something entirely unexpected — feel free to reach out!
Email:yan-jun.chen@connect.polyu.hk
WeChat:xzqm13143609845