UNDERGRADUATE RESEARCHER · NANJING UNIVERSITY

Minxin Dai

I am interested in efficient vision-language models and intelligent agent systems.

View publication GitHub ↗ Email ↗

UniversityNanjing University

MajorIntelligent Science & Technology

GPA4.64 / 5.00

Rank5 / 151

01 / About

Building efficient multimodal models and capable agents.

I'm an undergraduate student in Intelligent Science and Technology at Nanjing University. My research interest lies in efficient vision-language models that retain essential multimodal evidence under limited computational budgets.

I'm also interested in intelligent agents: how they plan, reason, use tools, and interact reliably with complex environments.

EDUCATION

Nanjing University

B.Eng. in Intelligent Science and Technology

National Scholarship Recipient

GPA 4.64 / 5.00 · Rank 5 / 151

02 / Publication

Selected publication

ICML 2026 · Accepted Co-first Author

UHR-BAT: Budget-Aware Token Compression Vision-Language Model for Ultra-High-Resolution Remote Sensing

UHR-BAT addresses token explosion and small-object challenges in billion-pixel remote-sensing imagery. It combines query-guided multi-scale input with a Region-wise Preserve-and-Merge strategy to retain salient local evidence under strict token budgets.

Text-derived global priors guide multi-scale visual input.
Redundant background regions are merged into compact token representatives.
State-of-the-art performance on XLRS-Bench, RSHR-Bench, and MME-RealWorld-RS.

View paper on ICML ↗

03 / Projects

Selected projects

Olympic Medal Prediction & Coach Effect Analysis

Nov 2024 — Feb 2025

Built a two-level random forest from athlete-level signals to country-level forecasts, achieving R² = 0.83 with uncertainty intervals. Combined mutual information, Cox survival analysis, and DID + DTW to study first-medal probability and elite coaching impact.

Random ForestCausal AnalysisMCM Honorable Mention

Lightweight LLM Deployment on Mobile Devices

Jan — Jun 2025

Built an end-to-end path from model compression to mobile inference. Applied group-wise 4-bit weight-only quantization to Qwen3-1.7B for approximately 4× compression, then integrated it with llama.cpp for real-time local generation.

Qwen3-1.7B4-bit Quantizationllama.cpp

04 / Recognition

Honors & awards

2025National Scholarship

National-level academic recognition

2024First-Class People's Scholarship

Top 10% · Nanjing University

2024 · 2025Outstanding Student

Nanjing University

2025CCPC Girls Contest

Bronze Medal

2024CCF Algorithm Competition

Third Prize

05 / Skills

Technical toolkit

Programming

Python, C++, C

Research

Efficient Vision-Language Models, Intelligent Agents

Tools

PyTorch, llama.cpp, Quantization

English

IELTS 7.0 · CET-6 604 · CET-4 615

GET IN TOUCH

Interested in efficient and capable AI?

Send an email ↗