Independent researcher
I work on language models, multimodal systems, and agentic learning.
My recent work focuses on how models reason, use tools, learn across modalities, and improve through post-training and interaction with environments.
About
I am an individual researcher working on large language models, multimodal systems, and agentic learning.
I care about a simple question: how do we move from models that generate convincing outputs to systems that can reason, act, and improve through interaction with the world? That question sits behind my interest in long-horizon planning, tool use, multimodal representation, and training methods that make model behavior more grounded.
My recent work spans agentic coding, multimodal modeling, reinforcement-learning methods, and efficient attention design. Across these projects, I keep returning to the same underlying theme: useful intelligence depends not only on model scale, but also on how architectures, environments, and training pipelines are designed together.
If you want a quick overview, start with the selected papers below or the essays in the blog.
Research Themes
I care about agents that can decide when to think, when to call tools, and how to revise a plan after seeing feedback from an environment. The core shift is from static reasoning to systems that sustain useful action over longer horizons.
Multimodal modeling matters when a system needs to read, inspect, and respond to the world rather than stay inside text alone. I am interested in model families that share representations across modalities and make interaction more grounded.
Better model behavior depends on more than prompts or scale. I am interested in reinforcement-learning algorithms, attention mechanisms, and infrastructure choices that make large-scale post-training stable, efficient, and aligned with real capability gains.
Selected Work
System-level report on a Qwen generation shaped by large-scale language modeling and post-training.
Report on a multimodal Qwen model family designed to operate across language and vision.
Work on coding agents that interact with tools and external environments instead of generating code in isolation.
Research on sequence-level policy optimization for reinforcement-learning settings in large model training.
Study of attention designs aimed at non-linearity, sparsity, and attention-sink-free behavior in language models.
Report on a Qwen image-generation effort, extending the model family beyond text and toward visual synthesis.
Writing
How the AI field is transitioning from static reasoning models to agentic systems that think in order to act.
Read the essayContact
If your interests overlap with agentic systems, multimodal learning, reinforcement-learning infrastructure, or the broader question of how foundation models become more grounded and useful, feel free to reach out.