cv

Ph.D.

EPFL (Swiss Federal Institute of Technology in Lausanne)

Machine Learning

Master

EPFL (Swiss Federal Institute of Technology in Lausanne)

Data Science

Master

University of California, San Diego

Computer Science

Bachelor

University of Electronic Science and Technology of China

Microelectronic Science and Engineering

Pre-training Algorithm Intern

High-Flyer | IQuestLab

Working on engineering and algorithmic research for next-generation LLM pre-training, including gradient-based influence data selection, SAE-guided diversity optimization, scaling laws, and low-quality data detection.

Research Intern

Oracle Labs

Developed a time series model that detects anomalous Linux sessions in cloud servers and used PySpark to process large-scale cloud data.

Algorithm Engineering Intern

DiDi (China's largest taxi-hailing platform)

Developed an algorithm for learning road segment weights from historical ride data, significantly improving route planning quality for ride-hailing services in production environment.

PriFT: Prior-Support Guided Supervised Fine-Tuning

Proposed a token-reweighting framework that uses support from a frozen pre-trained model to avoid the self-reinforcing dynamics of online reweighting. PriFT improves SFT across mathematical reasoning, code generation, and medical question answering while preserving exploration and providing a stronger initialization for subsequent RL.

LoRIF: Scalable Training Data Attribution for Large Language Models

Developed a novel, highly scalable method for training data attribution in large-scale models by exploiting the low-rank properties of gradients, cutting storage cost and query latency 20x. Enabled, for the first time, the ability to efficiently trace the output of a 70-billion-parameter LLM back to individual examples in their SFT training data.

Learning to Weight Parameters for Training Data Attribution

Identified the heterogeneity of attribution signal across parameters/layers in diffusion models and LLMs. Proposed a method to re-weight layers, boosting attribution accuracy and enabling interpretable attribution.

LLM Development from Scratch

Collaborative project with 25 PhD students to build a large language model from scratch. Engineered the pre-training pipeline, including environment setup and investigating optimal data mixing recipes for the training corpus. Implemented and validated the evaluation suite by reproducing the SmolLM2 benchmark to establish a robust performance baseline.

Enhancing Text-to-Image Generation with Reliable Random Seeds

Identified the significant role of initial noise in text-to-image inconsistencies for diffusion models. Proposed a method that identifies reliable random seeds to improve text-to-image generation, leveraging reliable seeds to synthesize high-quality data for fine-tuning diffusion models.

Controlling the Fidelity and Diversity of Deep Generative Models

Proposed an approach to bias generative models towards generating data with either enhanced fidelity or increased diversity. Enabled model training with data of better fidelity or diversity.

Interlock-Free Multi-Aspect Rationalization for Text Classification

Proposed a multi-stage training method to alleviate the interlocking issue in training interpretable models.

Project Supervisor

Improving Waste Detection and Sorting

Head Teaching Assistant

CS-233: Introduction to Machine Learning

Teaching Assistant

CS-401: Applied Data Analysis

Teaching Assistant

CS-233: Introduction to Machine Learning

Teaching Assistant

COM-407: TCP/IP Networking

Teaching Assistant

COM-112: Object-Oriented Programming (in C++)

Student Teaching Assistant

CS-456: Deep Reinforcement Learning

National Scholarship

China Collegiate Programming Contest - Gold Medal

China Collegiate Computing Contest - First Prize

First-class People's Scholarship

ACM International Collegiate Programming Contest (Asia Regional) - Bronze Medal

China Collegiate Computing Contest (Group Programming Ladder Tournament) - First Prize

First-class People's Scholarship

PriFT: Prior-Support Guided Supervised Fine-Tuning

arXiv (under review at NeurIPS 2026)

LoRIF: Low-Rank Influence Functions for Scalable Training Data Attribution

arXiv (under review at NeurIPS 2026)

Learning to Weight Parameters for Training Data Attribution

International Conference on Learning Representations (ICLR 2026)

Enhancing Compositional Text-to-Image Generation with Reliable Random Seeds

International Conference on Learning Representations (ICLR 2025 Spotlight)

Controlling the Fidelity and Diversity of Deep Generative Models via Pseudo Density

Transactions on Machine Learning Research

Interlock-Free Multi-Aspect Rationalization for Text Classification

arXiv