Howdy! I’m Weijia Zhang

Research

My research interests center on LLM agents, especially next-generation AI agents that bridge virtual and physical worlds through socially intelligent, tool-agnostic, and ethically grounded architectures.

• Multimodal agents: memory, reasoning, tool use, and multi-agent systems
• Conversational AI: anthropomorphism and social intelligence
• Post-training: agent SFT and RL

News

• 2026.06: 🎉 Joined TikTok as a Research Scientist Intern working on self-evolving agents! 🤖✨
• 2026.05: 🎉🎓 Graduated from UIUC in Math + Computer Science with the 2026 C.W. Gear Outstanding Undergraduate Student award! 🏆
• 2026.01: 🎉👑 Paper on bias inheritance was accepted to ACL 2026 as an Oral! 📄✨
• 2025.10: 🚀 SeeingEye was released as an arXiv preprint and is under review at EMNLP 2026! 👀📝
• 2025.07: 🎉 Joined Microsoft Research Asia as a research intern! 🔬💼

Gamedev

Beyond research, I am a passoinate indie game developer, feel free to check my game work on the game page. I am also willing to discuss the future of AI X Game.

Education

Yale University

M.S. in Computer Science

2026-09 - 2028-05

Thesis Track with Full Scholarship

GPA: 4.0/4.0

University of Illinois Urbana-Champaign

B.S. in Computer Science and Mathematics

2022-08 - 2026-05

2026 C.W. Gear Outstanding Undergraduate Student, one of two annual recipients; 2025 Dean's List

GPA: 3.7/4.0

Work Experience

TikTok

Research Scientist Intern, RSI / Self-Evolving Agents

Jun 2026 - Present

Project lead for TikTok's agent self-evolution framework, built from 0 to 1.

• Designed and shipped a closed Solver-Reflector-Evolver loop in which the Solver executes tasks, the Reflector attributes failures, and the Evolver iterates reusable skills; lifted F1 from 62.7% to 81.3% on unoriginal-video detection (9,600 real short videos) and from 76.5% to 87.3% on reposting-account detection (200 accounts).
• Designed a transferable Online Failure Discovery process: the Reflector mines failure patterns from scenarios and Solver traces, maintains a two-level scenario-to-failure-mode taxonomy, and distills high-value error clusters into reusable skills, so the same framework self-evolves across scenarios without human-predefined error types.
• Trained a video-understanding model: built an account-level SFT dataset of 74K accounts and 1.33M videos, ran distributed full-parameter SFT of Qwen3-VL 2B/4B/8B on 32 H100s with DeepSpeed, and calibrated thresholds across 53 checkpoints; the final 4B model cut false positives by 42.8% (318 to 182) against the 8B baseline at ~70% recall.

Microsoft

Research Intern, Large Language Models

Jul 2025 - Sep 2025

Built the Excel Coding Agent data engine and post-training pipeline for Excel Copilot's code agent.

• Produced execution-verified SFT/RL data by mining 20,000+ real-world workbooks, back-translating user queries from spreadsheet context, sampling multi-turn Office.js rollouts (write, execute, repair), and using SheetEngine final-state assertions as a programmatic verifier to filter incorrect and reward-hacking solutions.
• Post-trained the coding agent on this corpus with rejection-sampling SFT and RL over same-task pass/fail rollouts under execution-based reward, improving Office Scripts pass@1 by 15% against the production baseline on a held-out set.

Reborn Network

AI Agent Engineer

May 2023 - Jul 2023

Built real-time embodied role-playing agents in Unity VR.

• Built an embodied role-playing agent in a Unity VR environment, closing a real-time perception-dialogue-action loop across text, voice, and full-body VR actions at sub-second end-to-end latency.
• Introduced RAG/vector databases and dual-level (episodic + semantic) memory, improving cross-session recall accuracy from 38% to 61% on an internal multi-session dialogue eval.