Yuhan Ma

M.S. student in Computer Science at New York University

I am a graduate student in Computer Science at New York University. My interests are centered on recommender systems, generative image editing, and machine learning systems that are rigorous, reproducible, and useful beyond a single benchmark.

During my summer internship at Xiaohongshu (Rednote) in Shanghai, I worked on anime-domain post-training for FireRed-Image-Edit, spanning Flow Matching SFT, LoRA, reward-based optimization, and specialist-teacher distillation. I also built reciprocal ranking and constrained matching for the Darwin smart social matching project. My broader work includes retrieval-ranking pipelines and language model training.

Prior to NYU, I received my B.E. in Computer Science and Technology from Anhui University, where I graduated as an Outstanding Undergraduate Graduate of the Class of 2025. I also joined an exchange program at Deakin University.

I am seeking machine learning and recommendation roles for a spring internship (January-May 2027) or full-time work starting June 2027, and I am open to relocation nationwide. You can reach me by ym3447@nyu.edu or browse my GitHub.

I am currently interested in:

  • Recommender Systems: retrieval, ranking, re-ranking, user modeling, reciprocal matching, and evaluation design.
  • Generative Models and Post-Training: image editing, supervised fine-tuning, LoRA, reward design, and multi-teacher knowledge distillation.
  • ML Systems: reproducible data pipelines, checkpointing, experiment tracking, and low-friction deployment.

news

Sep 2026 Completed my summer internship at Xiaohongshu (Rednote) in Shanghai, working on anime-domain post-training and specialist-teacher distillation for FireRed-Image-Edit, and the Darwin social matching project through its Extra Mile Internal Venture Program.
Apr 14, 2026 Launched this academic-style personal website.
Mar 2026 Completed a reinforcement learning pipeline for LLM reasoning using SFT and GRPO on Qwen2.5-Math-1.5B.
Mar 2026 Finished a four-stage recommendation system for short-video feed ranking on KuaiRand-Pure.
Feb 2026 Built a GPT-style transformer language model from scratch with a streaming tokenizer and memmap training pipeline.
Aug 2025 Received overseas-study support from Anhui University for pursuing graduate study at New York University.
Mar 2024 Completed a machine learning engineering internship at National Quantum, working on image segmentation and deployment.
Jun 2023 - Nov 2023 Served as the student lead of the Information Technology Honors Class visiting cohort from the School of Computer Science and Technology at Anhui University, led the exchange cohort at Deakin University in Australia.

selected projects

  1. 01 Matching

    Darwin Project — Smart Social Matching Device

    Xiaohongshu (Rednote) · Extra Mile Internal Venture Program · Jul 2026 - Sep 2026

    Trained an intent-aware reciprocal ranker using multi-interest text embeddings and bidirectional demand-supply features. Built a reproducible feedback-to-training pipeline with 1,000+ directional feedback records across 10 events, versioned feature snapshots, missing-label masking, and temporal event holdouts.

    Developed a global matching engine with participant availability, repeat-pair exclusion, and bounded fairness adjustments, and verified small-instance solutions against an exact dynamic-programming baseline.

    Highlights: reduced held-out-event LogLoss by 10% versus a calibrated content-similarity baseline and improved total predicted matching utility by 10% over greedy pairing in event replays.

    Reciprocal Ranking Text Embeddings Constrained Matching Cold Start
  2. 02 RecSys

    Four-Stage Recommendation System for Short-Video Feed

    Jan 2026 - Mar 2026

    Built a full recommendation system on KuaiRand-Pure with multi-channel retrieval, coarse ranking, fine ranking, and rule-based re-ranking. The pipeline used time-based splitting and materialized intermediate candidates to keep experiments reproducible.

    Highlights: Recall@100 of 0.2476 / 0.2343 on validation and test, coarse-ranking validation AUC of 0.8149, and top-100 absolute Recall of 0.3523 / 0.3317 after reducing candidates from 500 to 100 per user. DIN-style multi-task fine ranking achieved test finish AUC of 0.6663, like AUC of 0.6219, and NDCG@20 of 0.0359. Re-ranking increased average unique tags per user by 0.302 and reduced the adjacent same-tag rate by 0.0171, with slight gains in NDCG@20 and Recall@20.

    ItemCF Two-Tower Graph Embedding LightGBM DIN
  3. 03 Benchmark

    Retrieval-Ranking Benchmark for Recommender Systems

    Sep 2025 - Nov 2025

    Built a unified recommendation benchmark on MovieLens-32M with negative sampling, Recall, and NDCG evaluation, and compared BPR-MF, GRU4Rec, SASRec, BERT4Rec, and dual-tower retrieval.

    Highlights: BERT4Rec achieved Recall@10 = 0.97 and NDCG@10 = 0.80.

    BPR-MF GRU4Rec SASRec BERT4Rec Dual-Tower
  4. 04 LLM

    GPT-Style Transformer Language Model from Scratch

    Jan 2026 - Feb 2026

    Implemented a byte-level BPE tokenizer and core transformer modules including causal self-attention, RoPE, RMSNorm, and SwiGLU, then paired them with a streaming tokenization and uint16 memmap pipeline.

    Highlights: processed 541M training tokens and 5.46M validation tokens, with a full-dev loss of 1.475.

    Transformer RoPE RMSNorm SwiGLU TinyStories
  5. 05 RL

    Reinforcement Learning for LLM Reasoning

    Feb 2026 - Mar 2026

    Built a training and evaluation pipeline for reasoning models based on Qwen2.5-Math-1.5B, covering zero-shot baselines, SFT, GRPO, reward design, automatic grading, and rollout-based updates.

    Highlights: improved Countdown validation accuracy to 32.4%.

    SFT REINFORCE GRPO Reward Design
  6. 06 Systems

    LLM Systems Optimization

    PyTorch + Triton

    Implemented a FlashAttention-2 optimized attention stack for a decoder-only Transformer, including a custom torch.autograd.Function reference path and a Triton fused forward kernel with causal masking.

    Highlights: built a GPU benchmarking and profiling workflow with CUDA-synchronized timing, torch.compile, mixed precision, Nsight profiling, and PyTorch memory snapshots to compare baseline attention against FlashAttention across model sizes and sequence lengths.

    PyTorch Triton CUDA FlashAttention torch.compile

experience

Xiaohongshu (Rednote) — FireRed-Image-Edit

Summer Intern · Anime-Domain Post-Training & Multi-Teacher Distillation · Jun 2026 - Sep 2026 · Shanghai, China

Fine-tuned FireRed-Image-Edit's MMDiT with Flow Matching SFT and LoRA for anime hair-color and accessory edits, improving held-out editing pass rate from 64.0% to 76.0%. Curated 2,000 human-reviewed editing triplets with pair-level deduplication and source-disjoint splits, including occluded hair and multi-character scenes.

Applied multi-task Flow-GRPO with task-aware rewards for instruction following, content preservation, and style consistency. Improved held-out editing pass rate by 6.5 percentage points and reduced unintended-edit rate from 18.0% to 9.0% versus the SFT baseline.

Delivered the anime-specialist teacher for team-led DanceOPD distillation. The consolidated student exceeded weighted LoRA merging by 6.0 percentage points in domain-averaged pass rate, while limiting anime-domain regression to 1.0 percentage point relative to the specialist.

Built reproducible checkpoint evaluation and cross-domain regression checks, including 200 blinded pairwise comparisons under matched sampling settings that verified a 60% human preference win rate over the SFT baseline.

National Quantum

Machine Learning Engineer Intern · Jan 2024 - Mar 2024

Worked on magnetic-particle image segmentation under limited-data conditions. I designed data processing and annotation workflows, used SAM for bootstrapped labeling, trained UNet variants in PyTorch, and improved validation performance with augmentation, Optuna, Focal Loss, and Dice Loss.

I also completed model export and production-side inference with ONNX Runtime, which gave me end-to-end exposure from training to deployment.

Education

New York University · Sep 2025 - May 2027 (expected)

M.S. in Computer Science, Courant Institute · GPA 3.9 / 4.0

Anhui University · Sep 2021 - Jul 2025

B.E. in Computer Science · Average Score 91 / 100 · Outstanding Graduate (2025) · First-Class Scholarship · Exchange Team Lead, Deakin University, Australia (Jun 2023 - Nov 2023)

Alongside academics, I served as class monitor of the Information Technology Honors Class, headed the college Arts Department, and captained the college soccer team.