Huiyao Zhang
🎓 Master's Student at UCAS · 🧠 Multimodal Reasoning · 🔬 Unified Vision Understanding & Generation
Hi! I am Huiyao Zhang (张晖耀), a Master's student in Computer Technology at the University of Chinese Academy of Sciences (UCAS), supervised by Prof. Ye Li. I collaborate with THUNLP & QIYUAN Lab, Tsinghua University, mentored by Zonghao Guo. Previously, I interned at TeleAI's Xingchen AGI Lab.
My research focuses on multimodal reasoning and unified visual understanding and generation. I am interested in how models use visual evidence to make well-supported decisions, and how they maintain consistent object grounding across understanding, segmentation, and image editing.
Detailed Profile
My current research explores fine-grained visual understanding and controllable editing in unified multimodal models. I study how models can consistently ground answers, segmentation masks, and image edits in the same intended object in the source image, while accessing the semantic and spatial details required by each task.
My previous work on CROSS-Bench and RIVET examines how visual evidence is associated with answer candidates and how those associations affect model decisions. I have also studied the adaptation of prompt-conditioned segmenters and geometry-aware visual modeling in IP-SAM and FGOS-Net, both accepted to ECCV 2026.
For research discussions or shared interests, feel free to email me at zhanghuiyao25@mails.ucas.ac.cn or connect on WeChat.
Connect on WeChat
Happy to connect over research, ideas, or shared interests. Please include a brief introduction when adding me.
Open full-size QR code ↗News
- 10/2026🎓 I was awarded the 2026 National Scholarship for Master's Students!
- 10/2026🎉 We released CROSS-Bench and RIVET for candidate-bound visual evidence use.
- 09/2026🔬 I joined THUNLP & QIYUAN Lab, Tsinghua University for research collaboration on unified visual understanding and generation.
- 07/2026🎉 IP-SAM and FGOS-Net were accepted to ECCV 2026!
- 03/2026🧑💻 I joined the Visual Understanding R&D Center at TeleAI's Xingchen AGI Lab as an intern, focusing on multimodal reasoning and MLLM post-training.
- 03/2026🎉 We released FGOS-Net for thin-structure SSMs.
- 01/2026🎉 We released Fluxamba for geological lineament segmentation.
- 08/2025🎉 We released ArgusCogito, a multimodal chain-of-thought framework.
- 01/2024🍇 YOLOv8-GP was published in Computers and Electronics in Agriculture (SCI Q1).
Education
-
—
University of Chinese Academy of Sciences (UCAS)M.S. Student in Computer Technology, supervised by Prof. Ye LiNational Scholarship for Master's Students, 2026Research in multimodal reasoning, foundation-model adaptation, and unified visual understanding and generation. -
—
Guangxi UniversityB.Eng. in Mechatronic Engineering, supervised by Prof. Jiqing ChenBackground in mechatronic systems, computer vision, and robotics.
Research Experience
-
—Present
THUNLP & QIYUAN Lab, Tsinghua UniversityResearch Collaboration · Mentored by Zonghao GuoResearching object grounding and spatial consistency across visual understanding, referring segmentation, and image editing, using BAGEL as the primary backbone; responsible for architecture design, implementation, training, and evaluation. This work is ongoing. -
—
TeleAI, China Telecom · Xingchen AGI LabIntern, Visual Understanding R&D Center · Mentored by Chuyi WangWorked on multimodal reasoning and post-training. Developed CROSS-Bench evaluation protocols and RIVET to study whether visual evidence affects the candidate it supports, separating evidence binding from decision strength. -
—Present
Technology and Engineering Center for Space Utilization, CASGraduate Student Researcher · Supervised by Prof. Ye LiResearch on prompt-conditioned foundation-model adaptation, cross-modal reasoning, and geometry-aware state-space models for camouflaged, medical, and remote-sensing image segmentation.
Selected Publications
[My Google Scholar]-
Visual Evidence Under Cross-Examination: Evaluating and Controlling Decision-Level Evidence Use in Vision-Language Models
arXiv preprint, 2026 · First author · Under reviewAccurate observations can still influence the wrong answer candidate. CROSS-Bench tests evidence validity and rebinding; RIVET preserves evidence identity and uncertainty while separating candidate-specific response from decision strength.
-
Bridging the Geometry Mismatch: Frequency-Aware Anisotropic Serialization for Thin-Structure SSMs
ECCV 2026 · Second authorRaster-style SSM scans propagate across rather than along thin structures. FGOS-Net couples frequency–geometry disentanglement with geometry-conditioned scanning to preserve directional continuity.
-
IP-SAM: Rethinking Prompt-Conditioned Segmentation for Prompt-Absent Deployment
ECCV 2026 · First authorIP-SAM restores SAM2's native prompt-conditioned decoding without user prompts, using intrinsic prompt generation and background gating; evaluations span camouflaged objects, salient objects, and medical polyps.
-
Fluxamba: Topology-Aware Anisotropic State Space Models for Geological Lineament Segmentation in Multi-Source Remote Sensing
arXiv preprint, 2026 · Second author · Under reviewRigid SSM scans fragment curved geological structures. Fluxamba uses topology-aware feature rectification and direction-sensitive context aggregation to preserve faint lineaments across planetary and terrestrial remote-sensing imagery.
-
ArgusCogito: Chain-of-Thought for Cross-Modal Synergy and Omnidirectional Reasoning in Camouflaged Object Segmentation
arXiv preprint, 2025 · Co-first author · Under reviewArgusCogito addresses target–background ambiguity through zero-shot cross-modal reasoning: scene-level conjecture, focused localization, and iterative mask refinement, evaluated on camouflaged-object and medical segmentation.
-
Efficient and lightweight grape and picking point synchronous detection model based on key point detection
Computers and Electronics in Agriculture · SCI Q1 · 2024Multi-step fruit detection and picking-point localization complicate robotic harvesting. YOLOv8-GP jointly predicts grape clusters and picking points in a lightweight keypoint-detection network.
Beyond Research
I really enjoy time to myself. Whether I am at my desk or at home, a pot of tea is often all I need to settle into something I enjoy for a long while. Sometimes that means studying, writing code, or thinking through my current research; other times, reading, watching a film, or catching a few episodes of anime. When a story really resonates with me, its characters and events tend to stay with me long after it ends.
I also love being active, especially running. In the evenings, I put on my headphones and run to music, feeling the breeze on my face as my breathing and footsteps find their rhythm. I love that feeling, and have always enjoyed the running scenes in Run with the Wind. Beyond running, I play basketball, badminton, table tennis, and football, and occasionally shoot pool with friends.
On weekends, I like going for walks with friends, without necessarily having a destination in mind. The Summer Palace is a favorite, especially on a clear day: a walk by the lake, a little sunshine, and a few photos of whatever catches my eye. Walking and talking with friends makes for the kind of unhurried afternoon I really enjoy.
Gaming is another part of my life. I take Honor of Kings fairly seriously and reach over 50 stars every season; I also enjoy Counter-Strike 2 and Wuthering Waves. Sometimes I play quietly on my own; other times, I team up with friends and get absorbed in a few matches.
This is the kind of everyday life I enjoy: quiet time on my own to do what I love, and, when I head out, the breeze, the sunshine, and friends.
My Hero · Lionel Messi
Lionel Messi has always been someone I deeply admire. I love his imagination on the pitch and his enduring dedication to football. His goals and achievements are extraordinary, but what I admire just as much is the person I see beyond the game: someone who cares about those around him, takes responsibility, and carries himself with kindness and warmth.
I was fortunate enough to visit Barcelona and explore FC Barcelona's stadium and museum. Seeing the shirts, trophies, and moments I had known for so long through photographs and matches was a special experience. Standing there, surrounded by those memories and achievements, made the stories I had followed feel much closer. It is a visit I will always treasure.
I hope to pursue excellence in my own field, just as he has in his. For me, that field is research. I want to stay curious, explore new ideas, and have the courage to take on difficult problems. Beyond publishing papers, I hope to do research that is genuinely creative, meaningful, and worth the effort it takes.
And beyond what I might achieve, I also care about the kind of person I become. I hope to remain kind and responsible, to treat the people around me well, and to be someone others can trust and feel comfortable around.
