About me
I am Peidong Wang (王培东), a direct-entry Ph.D. student in Computer Science at Northeastern University since September 2025, advised by Associate Prof. Shi Feng in the Data Mining group.
My research focuses on embodied AI, vision-language-action models (VLAs), and reinforcement learning for LLMs. I am currently a foundation-model research intern at XYZ AI Lab (行云至理研究院), working on Code Agent pre-training, RL, and training data. Previously, I interned at Microsoft Research Asia (MSRA) on embodied agents and LLM RL. I am mentored by Xufang Luo in both internships.
My work spans algorithm analysis, model training and data curation, and deployment on physical robots. I have first- or co-first-authored papers at ACL, EMNLP, and ACM Multimedia, with additional co-authored work in multimodal dialogue and reasoning.
Research experience · Publications · 中文简历 / Chinese CV
Research highlights
Embodied agents and online robot learning
SHAPER evolves skills and an execution harness through environment rollouts while keeping model weights and low-level executors frozen. It improves VLABench success from 28.25% to 34.50% and ESI-Bench accuracy from 32.5% to 49.8%. The code is available as an official Agent Lightning recipe.
On SO-ARM101, I work on VLA behavior cloning and online residual RL for tabletop manipulation. A π₀.₅ BC policy completed 39/39 trials across 13 placements in a medicine-box stacking task. For precise toy-sword placement onto a stand, two-stage residual Actor-Critic learning improved success from 2/12 (16.7%) to 9/12 (75.0%).
LLM RL and self-evolving systems
ACPO studies how off-policy degree changes importance-ratio clipping and token-level gradient dominance, then uses the analysis to design adaptive clipping for LLM reasoning. LANCE connects self-evolving data agents with iterative model improvement. I am also a co-author of AI4AI at Scale, a technical report on agentic system optimization.
Multimodal RL in practice
SAFE-QAQ combines audio-text reasoning, GRPO, and efficient inference for telecom fraud detection. Developed with China Mobile, the system is deployed in an anti-fraud production pipeline processing over 70,000 calls per day.
Recent updates
- 2026.08: SHAPER — our preprint on self-evolving embodied agents through skill-harness evolution, without updating model weights. Code is available as an official Microsoft Agent Lightning recipe.
- 2026.07: AI4AI at Scale — co-authored a technical report on bounded exploration and verification-gated agentic system optimization during my foundation-model internship at XYZ AI Lab.
- 2026.05: Joined XYZ AI Lab as a foundation-model research intern, working on Code Agent pre-training, RL, and training-data pipelines under the mentorship of Xufang Luo.
- 2026.04: SAFE-QAQ (ACL 2026, co-first author) — end-to-end audio-text fraud detection via RL, deployed with China Mobile in a production pipeline processing over 70,000 calls per day. [Code]
- 2025.09: Started my direct-entry Ph.D. in Computer Science at Northeastern University, advised by Associate Prof. Shi Feng in the Data Mining group.
- 2025.08: LANCE: Language Models as Continuous Self-Evolving Data Engineers (EMNLP 2025, first author) — language models that generate, clean, review, and annotate their own training data for iterative self-improvement. [Code]
- 2025.07: TeleAntiFraud-28k (ACM Multimedia 2025, co-first author) — an audio-text slow-thinking dataset and benchmark for telecom fraud detection. [Code & Data]
Earlier publications and competition results are listed on the Publications and Awards pages.