YJJiang, Yuhang
01 — About

JiangYuhang

Research Intern · Huawei Pisa Research Center
蒋雨航

I am a Research Intern at Huawei Pisa Research Center, working on multimodal agent memory, LLM post-training and automotive AI.

I recently completed an EIT Manufacturing dual master’s degree — a double degree from the University of Trento in Computer Science and SUPSI in Data Science. Before that, a bachelor’s in Engineering (Intelligent Science and Technology) from Anhui University.

I describe my research style as insight-driven and compute-efficient. I am a Kaggle Competitions Expert, and I care about how AI can be applied across many different fields. Besides research, I also like art, music, folklore and mythology very much.

Portrait of Yuhang Jiang
Plate 01 — Pisa, 2026
Appointment
Research Intern, Huawei Pisa Research Center — Apr to Oct 2026
Master’s
EIT Manufacturing Dual Master UNITN MSc, Computer Science — awarded Mar 2026 SUPSI MSE, Data Science — thesis defended Mar 2026, diploma Dec 2026
Bachelor’s
BEng, Intelligent Science and Technology Anhui University
Located
Pisa, Italy
Email
jyhtjtj@gmail.com
Elsewhere
Scholar   GitHub   LinkedIn   Kaggle
Curriculum
Brief, 2pp   Academic, 3pp
02

Selected Work

03 of 08
Implicit region-text matching picks the wrong aircraft while GeoSelect executes a spatial program and returns the correct instance

IEEE TGRS 2026Accepted

GeoSelect

Recast referring segmentation as executing a typed spatial program instead of matching embeddings. Training-free, single GPU, and more than twice the best prior training-free mIoU on RRSIS-D.

Project arXiv Code

Transfer attack success rate rising for standard surrogates and falling for robust ones as Input-Diversity probability increases

TMLR 2026Accepted as is

The Scissors Effect

Input Diversity is a near-default ingredient of transfer attacks — and on adversarially trained surrogates it reverses sign, costing a robust source 10.3 points of attack success. Leaving it on by default overstates the robustness of the model being evaluated.

Project arXiv Code

Left: an agent reaches the goal vicinity but cannot tell which mug matches the description. Right: Active Instance Verification resolves the attributes one at a time over a six-sector viewpoint topology

CVPR 2026FMEA Workshop · Poster

PInVerify

Reaching the goal vicinity is not finding the right instance. An offline embodied benchmark for active instance verification — 3,000 episodes across 18 categories, with trap views and unreachable sectors built into the navigation topology.

Project arXiv Code

All 08 publications →

03

News

08 entries
2026 — 09
GeoSelect is accepted at IEEE Transactions on Geoscience and Remote Sensing.
Accepted
2026 — 09
Serving as a reviewer for the InterpScience and TTCL workshops at NeurIPS 2026.
Service
2026 — 08
The Scissors Effect is accepted as is at TMLR.
Accepted
2026 — 08
A Circuit, Not The Circuit is updated on arXiv and recommended for Findings by ARR meta-review.
Preprint
2026 — 08
GeoSelect revision submitted to IEEE TGRS.
Under review
2026 — 07
GeoSelect submitted to IEEE TGRS.
Submitted
2026 — 05
Serving as a reviewer for the FMEA workshop at CVPR 2026.
Service
2026 — 04
Started as a Research Intern at Huawei Pisa Research Center.
Appointment
04

Research Interests

06 units
R.01

Agent memory

R.02

AI agents & reasoning

R.03

Mechanistic interpretability

R.04

Embodied AI

R.05

Vision-language models

R.06

Remote sensing