Linrui Ma
MIT Student • ML/AI Researcher • IChO Gold Medalist
I'm curious about the chemistry of intelligence, and the intelligence behind chemistry.
Undergraduate at MIT double majoring in AI & Decision Making and Chemistry (GPA: 5.0/5.0). My work spans deep learning for biology, computer vision, and AI for science: from cellular perturbation modeling to pixel-space image editing. IChO gold medalist and captain of China's national team.
About Me
I'm Linrui Ma, an undergraduate at MIT pursuing a double major in Artificial Intelligence & Decision Making and Chemistry (GPA: 5.0/5.0). My interests sit at the intersection of AI and the sciences: From cutting-edge deep learning models and the mechanisms and principles behind them, to the applications of machine learning algorithms and architectures in chemical and biological sciences.
This past summer I was an SDE Intern at Amazon on AWS Transform, building model-comparison infrastructure for an LLM-based migration orchestrator. In research, I work with Prof. Regina Barzilay at MIT CSAIL on cell-state prediction and specificity-aware generative models for selective protein binder design, and previously with Prof. Kaiming He on pixel-space generative models (MiniT2I / image editing, LoRA adaptation, and GCS dataset pipelines). Recent papers include Specificity-Aware Diffusion Steering (NeurIPS 2026), VGI-Bench (EMNLP 2026 main), and ProtQueSt (ICML 2026 workshops).
Before MIT, I completed my freshman year at Tsinghua University's Institute for Interdisciplinary Information Sciences (IIIS, Yao Class), studying computer science and AI. I was the captain of China's team and gold medalist (12th globally, 2nd in theory) at the 56th International Chemistry Olympiad, and a national gold medalist (2nd place) at the 37th Chinese Chemistry Olympiad.
Research Interests & Skills
Research Interests
Technical Skills
Publications
Specificity-Aware Diffusion Steering via Variance-Reduced Sequential Monte Carlo
NeurIPS 2026
Luran Wang*, Linrui Ma*, Hannes Stärk, Regina Barzilay
* Shared first authorship (arXiv listing pending update).
NeurIPS 2026 (Poster).
Inference-time diffusion steering that suppresses undesired modes while preserving the desired distribution via a specificity-aware target and variance-reduced Sequential Monte Carlo; evaluated on text-to-image generation and peptide-MHC binder design.
VGI-Bench: Probing Visual Intelligence in Video Generation Models
EMNLP 2026
Xuan He, Cong Wei, Yuhao Cheng, Linrui Ma, et al.
EMNLP 2026 (Main Conference). Main contributor (4th author).
Benchmark of 27 tasks and 810 instances for visual reasoning in video generation models, organized by a two-level taxonomy of task domains and skill tags. Current generative systems remain far from reliable: even Seedance 2.0 reaches only 51.0% under the evaluation criteria.
ProtQueSt: Query-Conditioned Retrieval-Augmented Generation for Protein Function Annotation
ICML 2026
Linrui Ma*, Yiwei Liang*, Yishu Yu**, Chuhan Joyce Qi**
* Shared first authorship. ** Shared second authorship.
ICML 2026 GenBio Workshop (Spotlight); ICML 2026 FM4LS Workshop (Poster).
First-authored RAG framework that pairs structure-aware retrieval with a query-conditioned contrastive retriever (FiLM + query-pooled negatives), reframing protein function annotation as query-dependent sequence understanding; Entity-BLEU 48.79 (+37% vs RAPM) on Prot-Inst-OOD.
MiniT2I: A Minimalist Baseline for Text-to-Image Generation
Technical Blog, 2026
Xianbang Wang†, Hanhong Zhao‡, Yiyang Lu‡, Kangyang Zhou, Linrui Ma, and Kaiming He§
(† project lead; ‡ core contributor; § supervisor)
Academic-scale pixel-space T2I baseline (MM-JiT). Contributed the LoRA downstream-adaptation extension (fine-tuning on small style datasets such as Naruto/Pokemon and producing the adaptation artifacts) and contributed code/notebooks to the public PyTorch release.
Featured Projects
NeuroMusic: EEG-Driven Therapeutic Music Generation
Link to projectBuilt an EEG-to-music pipeline that predicts valence/arousal (TSCeption), applies affective pivoting toward therapeutic targets, and synthesizes music with MusicGen/MusicLDM.
Multilingual Mathematical Reasoning in Neural Networks
Link to projectStudied how representation choice affects Transformer reasoning efficiency on multilingual math benchmarks; showed 5-10% token efficiency gains and attributed bottlenecks to training distribution via LoRA.
Labotex: AI-Powered Lab Report Generator
Link to projectBuilt an AI agent that turns experiment PDFs and CSV data into structured LaTeX lab reports using VLMs and chat models, with a Web UI for one-click PDF/LaTeX export.
High-Performance Matrix Multiplication with Multithreading & Distributed RPC
Implemented a high-performance matrix multiplication system in C with cache blocking, SIMD, multithreading, and cross-machine RPC; benchmarked for accuracy and speed.
Experience & Education
Software Development Engineer Intern
AWS Transform, Amazon
Boston, MA • Jun 2026 - Aug 2026
- ▸Built model-comparison infrastructure for the LLM-based orchestrator behind AWS Transform's VMware migration workflows
- ▸Designed automated evaluation scenarios and CloudWatch metrics across latency/cost, tool-calling, and reliability so model selection can be data-driven
- ▸Implemented the pipeline in Python and TypeScript (feature owned end-to-end under SDE I mentorship)
Undergraduate Researcher
MIT - Prof. Kaiming He's Group
Cambridge, MA • Feb 2026 - Jun 2026
- ▸Working on pixel-space generative models in the MiniT2I line under Prof. Kaiming He, including adapting the framework toward image editing and contributing LoRA-based downstream adaptation
- ▸Built a large-scale image-editing dataset pipeline to Google Cloud Storage
- ▸Diagnosed OOM failures on ~300GB tarballs and redesigned the flow with streaming shard extraction for reliable full-dataset upload
Undergraduate Researcher
MIT Computer Science and Artificial Intelligence Lab (CSAIL)
Cambridge, MA • Sep 2025 - Present
- ▸Collaborating with Ph.D. candidates under Prof. Regina Barzilay's supervision
- ▸Designing deep learning models to predict cell state changes in response to cellular perturbations
- ▸Developing probabilistic models for selective protein binder design
- ▸Applying AI techniques to advance medical and biological understanding
B.E. in AI & Decision Making + B.S. in Chemistry (Double Major)
Massachusetts Institute of Technology (MIT)
Cambridge, MA • Aug 2025 - Present
- ▸GPA: 5.0/5.0
- ▸Core Courses: Deep Learning; Algorithms; AI & Decision Making in Medicine; Biological Chemistry
- ▸Pursuing interdisciplinary education at the intersection of AI and chemistry
Undergraduate Student - Yao Class
Tsinghua University
Beijing, China • Aug 2024 - Jun 2025
- ▸GPA: 3.87/4.00
- ▸Majoring in Computer Science and Artificial Intelligence in prestigious Yao Class
- ▸Core Courses: Linear Algebra, C/C++ Programming, AI Principles & Techniques, Mathematics for CS & AI, Computer Systems, Introduction to LLM Applications
Key Awards & Achievements
First place in the MIT Informatics Tournament 2025-26 Winter Contest (Beginner's Round).
Top 1%, First Class Scholarship for Freshmen.
12th place globally, 2nd in theory. Captain of China's team.
2nd place nationally. Student representative speaker at closing ceremony.
Contact
I'm always willing to chat and discuss about both academic and non-academic topics.
Feel free to reach out!