Junyang Lin founded Pragmatik Labs to research agents for digital and physical environments. His earlier work with the Qwen team includes language-model training, multimodal systems and reinforcement-learning methods.[1][2][3]
Start with Qwen3 Technical Report →
GitHub 
X Career
Lin’s research spans models that use tools, learn across modalities and improve through interaction with environments. He announced that he had started Pragmatik Labs in Shanghai to develop next-generation agents across digital and physical settings. His work with Qwen provides a background in training and post-training large model systems.[1][2][3]
Contributions
One family for thinking and direct answers
Lin coauthored the Qwen3 technical report. The family brings reasoning-oriented and direct-response behavior into a shared model setup, with dense and mixture-of-experts variants. Its post-training pipeline addresses reasoning, instruction following and tool use, while supporting control over how much computation a response spends on thinking. The report describes a team effort across pretraining, post-training and evaluation.[3]
Optimize a response at sequence level
He also coauthored Group Sequence Policy Optimization. GSPO defines its importance ratio using the likelihood of an entire response and applies clipping and optimization at sequence level. The paper contrasts this with token-level ratios and evaluates training stability and efficiency, including mixture-of-experts models. The method changes how a sampled response contributes to the policy update, connecting the optimization unit more directly to the sequence-level reward.[4]
Add a gate after attention
The gated-attention paper studies applying a learned, nonlinear gate to attention outputs. It examines sparsity, training stability and attention sinks, where a model assigns disproportionate attention to a small set of tokens. Lin is a coauthor; the experiments investigate architectural choices rather than attributing an entire model’s behavior to a single contributor.[5]
Tools, writing & talks
Writing, research & talks
Start here
Qwen3 Technical Reportarxiv.orgA report on a model family that combines reasoning and direct responses with staged training.
Start with the model-family overview and the explanation of thinking versus direct responses. Basic familiarity with language models helps; read the training stages before comparing the benchmark tables.
Coauthor[3]- Group Sequence Policy Optimizationarxiv.org
A sequence-level reinforcement-learning method for stable language-model post-training.
Coauthor[4] An architectural study of gating, sparsity and attention-sink behavior.
Coauthor[5]