🚀🚀🚀 I'm on the job market for full-time Research Scientist / Engineer or related positions in Large Language Models (Post-Training & Agentic RL)! Please reach out to me if you are interested in my research!
Research Engineer / Scientist focused on LLM Post-Training & Agentic RL — building forkable agent runtimes, state-faithful SFT pipelines, verifiable container environments, and evaluation benchmarks.
🌐 Personal Homepage · 🎓 Google Scholar · 𝕏 X (@shatianmin20008) · 🤗 HuggingFace · 📫 Tianming.Sha@stonybrook.edu
- Octopus (Project Page · Tech Report · 🤗 Env & Corpus): GPU-resident execution DAG (
89.6%KV hit rate), hijack-free zero-copy KV mounts (21×lower hijack), and 1D packedDAG-GRPO— lifting Qwen3.5-4B (BM25) to 49.6% on BrowseComp-Plus (+17.5 pts), surpassing dense-retrieval agents up to 250× larger - XYZ-Aquila & AxisAgentic (AI4AI at Scale): State-faithful SFT, 7,000 open trajectories, and 2,000+ rootless
udockerRL environments trained across 2,000 H200s - Agent² RL-Bench & BenchScope: NeurIPS 2026 Evaluations & benchmark signal analysis (@ Microsoft Research Asia)
- SkillCenter / LangSkills & PaperFarm: 119,608 source-grounded agent skills & unattended Git-worktree research loop
- ThinkPrior & FAST-CAD: Zero-rollout difficulty priors for RLVR & AAAI 2026 Oral