[2026.6] Appointed as Adjunct PhD Co-Supervisor at Shanghai Jiao Tong University, focusing on AI safety and trustworthy AI systems, advancing joint talent development and frontier research between the PuJiang National Lab and academia.

[2026.6] Appointed as Lead Principal Investigator (课题负责人) of the “Science Intelligence (AI4S) Secure Service Middleware” sub-project under China’s “New Generation AI” National Major Project, with total funding of ¥10 million, building security infrastructure for scientific computing, drug discovery, and industry simulation.

[2026.5] Released the Safactory technical report, presenting a scalable agentic infrastructure for training trustworthy autonomous intelligence through a closed-loop pipeline spanning parallel simulation, trustworthy data, and autonomous evolution.

[2026.5] NaviMaster was accepted to ACL 2026 Main, unifying GUI and embodied navigation with a mixed-trajectory reinforcement learning framework for stronger cross-domain generalization.

[2026.4] Released Mokai, a full-stack agent security toolbox with 150+ tools spanning risk simulation, trusted data, and evolutionary defense for practical deployment.

[2026.4] Deliberative Searcher was accepted to ACL 2026 Oral, integrating certainty calibration and retrieval-based search into constrained reinforcement learning to improve reliability and calibration in open-domain QA.

[2026.4] From Coarse to Fine was accepted to ACL 2026 Findings, introducing the fine-grained WEval benchmark and WRL reward modeling framework for writing-centric generation with more precise requirement adherence.

[2026.3] Released DRIFT, later accepted to ACL 2026; the related paper proposes a dual-model framework that decouples knowledge extraction from reasoning through implicit fact tokens for efficient long-context inference, stronger robustness, and resistance to jailbreak prompts.

[2026.2] Released the AOT perceptual robustness project; the related paper introduces the AOT-SFT adversarial dataset and an attacker-defender self-play framework that continually improves multimodal perceptual robustness in visually complex scenes while reducing hallucinations.

[2026.2] TPRU was accepted to ICLR 2026 Oral, introducing a temporal and procedural understanding dataset plus an RL training pipeline that substantially improves lightweight multimodal models across robotics and GUI scenarios.

[2026.2] Released SafeVerse: a safe and trustworthy digital twin arena for embodied AI, turning ordinary videos into interactive, physics-aware 3D twin scenes within minutes, then extending them with attack-oriented editing and online agent evolution.

[2025.12] Released BioBridge: letting LLMs truly understand proteins without sacrificing general ability; the related paper was presented at BIBM 2025 (CCF-B), combining specialist protein reading with LLM reasoning to reach near-specialist performance on real biological tasks.

[2025.11] As the sole corresponding author, a LLM reward model based on uncertainty modeling was presented at EMNLP 2025 Main (Oral), which can significantly improve the content thinking quality of large model inference training.

[2025.9] Released a roadmap for Safe and Trustworthy Embodied AI with Xin Tan, Chaochao Lu, and Bowen Zhou, defining the field with ten core principles and an L1-L5 maturity model.

[2025.8] As the sole corresponding author, a distributed KV Cache architecture optimized for MoE models was released (PiKV: KV Cache Management System for Mixture of Experts).

[2025.8] Released the SOTA CoT-PRM model (VRPRM), achieving Test-time Scaling effect of Best-of-N approaching the theoretical limit value Pass@K, surpassing the SOTA model by 118% with only 1/8 of traning data.

[2025.7] As Core Lead, I led the knowledge enhancement and “Deliberation Search Mode” related modules of SafeWork-R1, with the results released at WAIC 2025.; the related paper presents the SafeLadder framework for jointly improving multimodal reasoning safety and general capability.

[2025.6] As the corresponding author, he guided the internship student to submit a paper, which was included in ICCV 2025. The relevant achievements have refreshed the SOTA of multi-modal retrieval, and it can provide the function of precise memory retrieval for embodied AI with 500,000 frames. demo

[2025.5] A comprehensive review article on AI tracing was published in Artificial Intelligence Review, spanning 60 pages and comprising 20,000 words.

[2024.10] Join Safety and Trustworthy AI Center in the Shanghai AI Laboratory, responsible for knowledge enhancement for LLMs.

[2023.10] First author paper, using a dynamic graph network evolution engine for accelerated simulation of complex transportation systems.

[2022.10] Join Shanghai AI Laboratory, mainly responsible for AI security evaluation systems and multi-agent simulation platforms. I earned an ‘Excellent’ performance rating for two consecutive years.

[2022.11] A paper is included in the 1st Learning on Graphs Conference, mainly studying the evolution issue of dynamic graphs.

[2022.3] Collaborative paper published in Nature Machine Intelligence.

[2021.9] The first author’s graph computing paper has been published in SIGMOD 2021 Oral, which is the result of an internship at Ant Group.

[2020.2] Life’s first paper was published in Knowledge-Based Systems