What If the Bug Had a Dial?
FLAS turns model behavior into a continuous dial, from fixing creature-metaphor bugs to controlled brand voice, all at inference time without retraining.
Read storyResearch / agents / safety
Xinjie Shen沈鑫杰
Toward capable, trustworthy agents.
As AI becomes embedded everywhere in everyday life, I study how intelligent systems can act safely, protect privacy, and remain reliable over long horizons. My research spans multi-turn attacks, timely defenses, and state-aware safety for tool-using agents; privacy in physical and interactive environments; and the evaluation and training of long-horizon agents. These directions come together in the research initiative Everywhere Safety.
I am a Ph.D. student in Machine Learning at Georgia Tech, advised by Prof. Pan Li. My work has been shaped by research experiences and collaborations across Georgia Tech, CMU, Qwen, Amazon Web Services, and Microsoft Research Asia.
01Writing & ideas
All postsIdeas, experiments, and notes from the research process.
02Recently
All news- Released SEAD, a state-based framework for attack and defense in tool-using agents. Project · Code and data.
- Released VHD-Play, a Qwen technical report on generating verifiable agentic RL environments from solved mechanisms. I also wrote about how we iterated on its teaser figure.
- Qwen3.8-Max-Preview reached No. 1 on NVIDIA’s FlashInfer-Bench during our kernel-optimization work. Our longer real-world self-improvement runs exceeded 60 hours and 10,000 tool calls. Read the report.
- Gave an invited talk at Tsinghua University PI Lab on physical-world privacy awareness in LLMs and VLM agents.
- Released E-Commerce Bench, a 365-day benchmark for long-horizon autonomous business operation.
- Gave the Jiangmen / TechBeat talk Distributed Hidden Malice: Attacks and Defenses, connecting CKA-Agent attacks with TurnGate defenses for multi-turn dialogue.
03Selected research
All publicationsarXiv preprint
SEAD: A State-Based Perspective on Attack and Defense in Tool-Using Agents
A state-based framework for attacks and defenses in tool-using agents, pairing execution-guided trajectory search with state-aware pre-execution defense.
Qwen Technical Report
Verifiable Hidden Dynamics Play: Generating Agentic RL Environments from Solved Mechanisms
Generate diverse, stateful agentic RL environments with verifiable rewards by deriving both dynamics and evaluation from solved mechanisms.
Qwen Technical Report
E-Commerce Bench: Evaluating LLM Agents on Long-Horizon Autonomous Business Operation
A reproducible 365-day benchmark for agents operating online stores under negotiation, delayed outcomes, market shocks, and persistent memory demands.
ICML 2026 AIWILD Workshop
One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue
A response-aware monitor trained to intervene at the first harm-enabling turn while preserving benign exploratory dialogue.
NeurIPS 2026
How Far Are VLMs from Privacy Awareness in the Physical World?
An interactive audiovisual benchmark showing that current VLMs struggle to turn privacy cues into situated behavior.
04Research in motion
YouTube channel ↗Short explainers on the ideas behind the papers. Watch on YouTube.
4:52SEAD
State-grounded attacks and defenses for tool-using agents.
4:10TurnGate
Why the timing of safety intervention matters.
4:47The Trojan Knowledge
When harmless pieces connect across a conversation.
4:49Physical privacy
From recognizing privacy to acting on it.
4:40VHD-Play
Scaling agentic RL environments with verifiable rewards.
5:00A year in business
Long-horizon business operation with E-Commerce Bench.
05Along the way
2026 — prev.
Qwen
Former Post-Training Research Intern, Long-Horizon Agents
Built agentic RL data and training pipelines for Qwen3.7-Max and Qwen3.8-Max. My work spanned verifiable environment generation, long-horizon evaluation, kernel optimization, and real-world self-improvement.
2025 — now
Georgia Institute of Technology
Ph.D. Student in Machine Learning
Advised by Prof. Pan Li. I work on trustworthy agentic intelligence, including distributed-intent attacks and defenses, physical-world privacy, and state-aware tool-use safety.
2024 — 2025
Amazon Web Services
Research Intern, Sparse Retrieval
Developed an $\ell_0$ sparsification method for inference-free sparse retrievers. The released model series averages 1.8 million downloads per month.
2024 — 2024
Microsoft Research Asia
Research Intern, Autonomous R&D Agents
Worked with Dr. Jiang Bian on data-centric automatic R&D and helped build RD-Agent. Received the Star of Tomorrow award.
Let’s talk.
Research conversations and collaborations are welcome.
xinjie@gatech.edu ↗/Book a conversation ↗