How Much Vision Does Multimodal Reasoning Need? Vision-Stripping for Multimodal Benchmarks
Weijia Zhang, Zijia Liu, Tianyi Zhang, Ruiqi Chen, Lian Zhang, Haoru Li, Haoqi Chen, and Jiaxuan You
Under review at NeurIPS 2026
CV
Education
- M.S. in Computer Science, Yale University, Sep 2026 - May 2028
- Thesis Track with Full Scholarship
- Incoming student
- B.S. in Computer Science and Mathematics, University of Illinois Urbana-Champaign, Aug 2022 - May 2026
- GPA: 3.7/4.0
- 2026 C.W. Gear Outstanding Undergraduate Student, one of two annual recipients
- 2025 Dean’s List
Work Experience
- Machine Learning Engineer Intern, RSI / Self-Evolving Agents, TikTok, Jun 2026 - Present
- As project lead, designed and shipped TikTok’s agent self-evolution framework from 0 to 1 with a closed Solver-Reflector-Evolver loop: 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 directly from scenarios and Solver agent traces, continuously updates a two-level “scenario to agent failure mode” taxonomy, and distills high-value error clusters into reusable skills, so the same framework self-evolves across scenarios without any 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 performed cross-domain evaluation and threshold calibration over 53 checkpoints. The final 4B model cut false positives by 42.8% (318 to 182) against the 8B baseline at ~70% recall.
- Research Intern, Large Language Models, Microsoft, Jul 2025 - Sep 2025
- Built the Excel Coding Agent data engine producing execution-verified SFT/RL data for Excel Copilot’s code agent: mined 20,000+ real-world workbooks, back-translated user queries from spreadsheet context, sampled multi-turn Office.js rollouts (write, execute, repair), and used 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.
- AI Agent Engineer, Reborn Network, May 2023 - Jul 2023
- 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.
Publications
200+ citations on Google Scholar.
Every Act Has Its Price: Compressed Moral Composition in Frontier LLMs
Weijia Zhang*, Ruiqi Chen*, Yunze Xiao*, and Weihao Xuan
arXiv:2606.11232
CUADebug: Diagnosing and Repairing Computer-Use Agent Failures
Weijia Zhang, Kunlun Zhu, Zeyi Liu, Yinting Chen, Tianyi Ma, Jiateng Liu, Jiaxun Zhang, Bingxuan Li, Pan Lu, Xiangru Tang, Heng Ji, and Jiaxuan You
EMNLP 2026
Understanding and Mitigating the Bias Inheritance in LLM-based Data Augmentation on Downstream Tasks
Miaomiao Li, Hao Chen, Yang Wang, Tingyuan Zhu, Weijia Zhang, Kaijie Zhu, Kam-Fai Wong, and Jindong Wang
ACL 2026, Oral
AgentDebugX: An Open-Source Toolkit for Failure Observability, Attribution, and Recovery in LLM Agents
Kunlun Zhu, Xuyan Ye, Zhiguang Han, Yuchen Zhao, Bingxuan Li, Weijia Zhang, Muxin Tian, Xiangru Tang, Pan Lu, James Zou, Jiaxuan You, and Heng Ji
EMNLP 2026 (System Demonstration)
SeeingEye: Agentic Information Flow Unlocks Multimodal Reasoning in Text-only LLMs
Weijia Zhang*, Zijia Liu*, Haoru Li*, Haoqi Chen*, and Jiaxuan You
arXiv:2510.25092 · AAAI 2026 under review
Where LLM Agents Fail and How They Can Learn From Failures
Kunlun Zhu, Zijia Liu, Bingxuan Li, Muxin Tian, Yingxuan Yang, Jiaxun Zhang, Pengrui Han, Qipeng Xie, Fuyang Cui, Weijia Zhang, Xiaoteng Ma, Xiaodong Yu, Gowtham Ramesh, Jialian Wu, Zicheng Liu, Pan Lu, James Zou, and Jiaxuan You
Failure Modes in Agentic AI Workshop, ICML 2026
Artificial Leviathan: Exploring Social Evolution of LLM Agents Through the Lens of Hobbesian Social Contract Theory
Gordon Dai*, Weijia Zhang*, Jinhan Li, Siqi Yang, Srihas Rao, Arthur Caetano, and Misha Sra
The First Workshop on AI Behavioral Science, ACM SIGKDD 2024
Project Experience
- OpenManus & OpenManus-RL, Core Author, Jan 2025 - Jul 2025
- Collaborated with the official OpenManus team as a core contributor to OpenManus-RL, a verl-based agent RL post-training framework; the ecosystem has accumulated 70,000+ GitHub stars.
- Connected SFT, RL (PPO/GRPO), interactive environments, and trajectory rollout into one unified pipeline to train ReAct agents and tool use on long-horizon tasks such as GAIA, WebShop, and ALFWorld.
- Unified multi-turn environment interaction, tool actions, validation, and reward under a single ReAct trajectory protocol, and released a dataset of ~49K high-quality agent trajectories spanning OS, Web, Database, Knowledge Graph, Household, and E-commerce scenarios.
- AgentDebugX, Project Lead, Aug 2026 - Present
- As co-lead, designed and implemented AgentDebugX, the first comprehensive debugging framework for long-horizon agent trajectories; its Detect-Attribute-Recover-Rerun loop turns LLM agent execution traces into auditable root-cause diagnoses, repair suggestions, and verifiable reruns across multi-agent, tool-using, and computer-use agents.
- Designed the core diagnostic agent DeepDebug, which localizes root causes over multiple rounds via global trajectory investigation; on the 184 Who&When trajectories it raises agent-and-step attribution accuracy from 21.7% to 28.8%, and on GAIA it repairs 13/73 failed tasks in a single round, lifting overall accuracy from 55.8% to 63.6%.
- Built an open-source infrastructure spanning a Python library, CLI, computer-use support, agentic skills, and an Error Hub, providing unified trajectory ingestion, failure reuse, and regression validation for LangGraph, CrewAI, OpenAI Agents SDK, OpenTelemetry, and OSWorld.
- SeeingEye, Co-first Author / Researcher, Oct 2025 - Present
- Proposed an agentic information-flow framework that enables text-only LLMs to perform multimodal reasoning through structured, tool-mediated textual context.
- Designed the perception-to-reasoning pipeline for multimodal tasks while keeping the backbone model text-only.
- Studied how tool use, intermediate representations, and memory-like context unlock multimodal reasoning behavior in text-only LLMs.
- Released the work as an arXiv preprint, currently under review at AAAI 2026.
- Paper: SeeingEye
Activities and Awards
- May 2026: C.W. Gear Outstanding Undergraduate Student, one of two annual recipients
- 2025: UIUC Dean’s List
- Aug 2025: RecSys 2025 Reviewer
- Sep 2022: UIUC MechMania 28 AI Hackathon, 2nd Place
Skills
- Programming languages: Python, C/C++, C#
- Frameworks and libraries: VERL, vLLM, LangGraph, LangChain, PyTorch, TensorFlow
- AI focus: RL/post-training, LLM agents, multimodal reasoning, self-evolving agents, SFT, RAG, evaluation