CircuIdea
Keep proposals, reviews, and resource contributions in one workspace. Start in your browser, sync to a private GitHub repository, and use signed Git deliveries when needed.
Tong Zhang Stone
I am a Ph.D. student in Computer Science at Fudan University, advised by Yun Peng and Tao Xie. Previously, I received my M.S. in Computer Engineering from the University of California, Irvine, advised by Xiaohui Xie, and my B.Eng. from the Southern University of Science and Technology, advised by Yu Zhang.
My research focuses on Coding Agent and Multimodal Reasoning. Feel free to get in touch!
I believe the most exciting frontiers for AI are curiosity and abstraction.
“What magical trick makes us intelligent? The power of intelligence stems from our vast diversity, not from any single, perfect principle.”
— Marvin Minsky, The Society of Mind
Shui!
Share persistent agent sessions, delegate work to teammates, and bring experiments, results, and figures back into one research workspace.
ClawPet
A voice-first embodied AI companion with simple physical expression, agent-to-agent chat, and a 3D-printable shell for desktop life.
MockMyInterview
An AI-powered mock interview experience for practicing technical and behavioral interviews with interactive feedback.
LaTeX Blog Template
A reusable LaTeX blog system template with blog-style landing pages, automated compilation, and GitHub Pages deployment.
LiveUIBench: Evaluating Agent as App Workflows
LiveUIBench evaluates agent-as-app orchestrations across changing goals, revisions, service actions, saved state, and return visits through a shared runtime.
When Does Recurrence Become an Algorithm? Convergence Selection in Weight-Tied Looped Transformers
When does a weight-tied looped transformer — one block applied T times — implement an actual algorithm? Identifies a budget law pricing the learned frontier’s speed by the training contract, and introduces a convergence-time instrument τ(n, i) that predicts out-of-distribution fate where standard metrics saturate.
GEB-Bench: Abstract Structures Told in Many Voices
A benchmark whose unit is an abstract structural motif — self-reference, a strange loop, a Möbius twist — each told in several voices: a natural scene, a folk story, a theorem, and a programmatic skeleton. Models recognize a structure within one voice far better than they carry it across voices. Fully generative and released with its pipeline.
Knowing When to Stop Thinking
EFEReasoning dissects how progress, disagreement, stopping, and online world-model updates affect test-time planning, showing that a better internal objective does not always produce a better action.
InfiniteWorld: Separating Visual Evidence from Familiar Physics
InfiniteWorld tests whether vision-language models follow observed videos when their physical rules depart from familiar Earth priors, revealing a persistent bias toward the world they already know.
When LLM Defenses Backfire: Characterizing Safety, Performance, and Cost Trade-offs
A systematic study of how jailbreak defenses trade off safety against task performance, over-refusal, and inference cost. Benchmarks 11 defenses across 6 open-source LLMs and 5 datasets, organizing them by operational strategy to expose their deployment-time side effects.
Beyond Pixels: Human-Readable SVG Generation
The first study to adapt a pretrained autoregressive vision-language model for image-to-SVG code generation, producing human-readable graphics from simple images.
Tong Zhang, Yuxuan Cheng, Ruimin Zhang, Yun Peng, and Tao Xie†, “LiveUIBench: Evaluating Agent as App Workflows”, Preprint, 2026.
Tong Zhang, Junhao Hu, Yun Peng, and Tao Xie†, “When Does Recurrence Become an Algorithm? Convergence Selection in Weight-Tied Looped Transformers”, Preprint, 2026.
Tong Zhang, Zhiyuan Shi, Yun Peng, and Tao Xie†, “GEB-Bench: Abstract Structures Told in Many Voices”, Preprint, 2026.
Tong Zhang, Zhiyuan Shi, Yun Peng, and Tao Xie†, “Knowing When to Stop Thinking: A Dissection of Expected Free Energy as a Test-Time Compute Controller for Latent World-Model Planning”, Preprint, 2026.
Tong Zhang, Yuxuan Cheng, and Yun Peng†, “InfiniteWorld: Separating Visual Evidence from Familiar Physics”, Preprint, 2026.
Tong Zhang, Zexin Li, Simin Chen, and Yun Peng†, “When LLM Defenses Backfire: Characterizing Safety, Performance, and Cost Trade-offs”, Preprint, 2026.
Tong Zhang, Haoyang Liu, Peiyan Zhang, Yuxuan Cheng, and Haohan Wang, “Beyond Pixels: Exploring Human-Readable SVG Generation for Simple Images with Vision Language Models”, Preprint, 2023.
Zebin Yun, Shengjie Niu, Tong Zhang, and Jian Huang, “Proof of Forgeability: Challenging the Evidentiary Reliability of Membership Inference”, Under Review, SaTML 2027.
Sinin Zhang, Yunfei Xie, Yuxuan Cheng, Haoyu Zhang, and Tong Zhang, “PhysNote: Self-Knowledge Notes for Evolvable Physical Reasoning in Vision-Language Model”, ES-Reasoning @ ICLR, 2026.
Haoyu Ma, Tong Zhang, Shanlin Sun, Xiangyi Yan, Kun Han, and Xiaohui Xie, “CVTHead: One-shot Controllable Head Avatar with Vertex-feature Transformer”, IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2024.
Bu Jin, Xinyu Liu, Yupeng Zheng, Pengfei Li, Hao Zhao, Tong Zhang, Yuhang Zheng, Guyue Zhou, and Jingjing Liu, “ADAPT: Action-aware Driving Caption Transformer”, IEEE International Conference on Robotics and Automation (ICRA), 2023.
Sep. 2025 – Dec. 2025
Research Intern
Developed a WeChat Mini Program agent based on vision-language models, and designed agent memory together with experience extraction mechanisms.
May 2025 – Aug. 2025
Research Intern
Proposed an energy-based attention refinement method for image editing, and explored generate–reflect–reason loops for multi-round controllable optimization.
Nov. 2024 – Apr. 2025
Research Intern
Studied defense mechanisms for large language models, their side effects, and token compression for audio language models.
May 2022 – Sep. 2022
Lightweight OCR Models for OpenCV
Implemented PP-OCRv2 detection in OpenCV Zoo, high-level C++ APIs, and evaluation metrics for text detection.
我是大水理论和大睡理论的信徒，希望以后能练成薄肌。
