Tianjun (Grady) Gu

Visiting Research Intern

NYU Center for Robotics and Embodied Intelligence (CREO) · AI4CE Lab

Advisor: Chen Feng · NYU Tandon School of Engineering

• Investigating spatial representation learning for MLLMs by testing whether complementary multi-view, panoramic, and egocentric observations yield more viewpoint-consistent scene representations and improve cross-view grounding and spatial reasoning.

Visiting Research Intern

Tsinghua University Spatial Intelligence Lab · Beijing, China

Advisor: Yiming Li

• First-authored Towards Spatial Supersensing in the Wild (ECCV 2026), introducing VSI-Super-Wild, a four-task benchmark for continual agent-object-environment world-state reasoning with 442 videos (284.52 hours) and 6,980 human-verified QA pairs; evaluated 13 MLLMs and found average performance drop from 35.0 on 0-10-minute videos to 26.3 beyond 120 minutes.
• Contributed to additional collaborative projects on generalist agents and foundation-model evaluation, including Towards Generalist Game Players: An Investigation of Foundation Models in the Game Multiverse.

Graduate Researcher

Digital Media & Computer Vision Lab (DMCV) · Shanghai, China

SJTU–ECNU Joint Lab · Advisors: Lizhuang Ma and Xin Tan

• Developed DORAEMON, a zero-shot navigation framework with hierarchical semantic-spatial and topological memory and safety-aware VLM planning; achieved state-of-the-art SR and SPL on HM3D, MP3D, and GOAT.
• Contributed to national-level, municipal, and industry-sponsored research programs on embodied intelligence, autonomous driving, agent evaluation and safety, and AI-enabled social simulation.

Foundation Model Research Intern

Meituan · Infrastructure R&D Platform / M17 · Shanghai, China

• Improved LongCat-Flash-3B's omni-modal capabilities, focusing on multilingual instruction following through targeted data construction, supervised post-training, and systematic evaluation.
• Built an end-to-end GUI-agent data engine for LongCat-Flash, converting web-sourced videos into interaction trajectories and training data across generation, quality control, and model post-training.
• Developed reinforcement-learning schemes for interactive multimodal models spanning audio-visual and spoken interaction, covering task construction, reward design, and multi-turn policy optimization; also explored interactive world models through benchmark and memory-oriented evaluation design.

Research Intern

Tencent Youtu Lab · Shanghai, China

• Led SESB (under review), an inference-time spatial memory with query-relevant belief access, targeted video retrieval, and local updates; achieved 87.1 on VSI-Bench, versus 50.8 with static belief and 59.9 under the same frame budget.
• Led the EscherVerse study (Nature Communications, Revision), building the dataset from 11,328 videos, with an 8,000-example benchmark and a 35,963-example instruction-tuning set.
• Explored autonomous research agents and automated research workflows, and supported post-training of Youtu foundation models through data construction, training-pipeline development, and evaluation.

AIGC Multi-Model Algorithm Intern

Baidu · MEG / Library Strategy Group · Shanghai, China

• Built scenario- and character-specific LoRA datasets and fine-tuned image-generation models, improving animated-character consistency by 38 percentage points.
• Developed automated multi-character LoRA training pipelines and supported the production launch of Baidu Wenku's one-sentence comic feature.

Machine Vision Algorithm Engineer

SAIC Anji Intelligent IoT Technology Co., Ltd. · Shanghai, China

• Built cross-site warehouse vision datasets using CenterNet-based candidate mining, similarity filtering, and end-to-end preprocessing pipelines.
• Trained and deployed workwear-detection models across multiple warehouse sites, iterating through on-site evaluation, error analysis, and operational feedback.