Profile
Education
Columbia University, The Fu Foundation School of Engineering and Applied Science
Sep 2023 – Dec 2025Courses: Machine Learning, Natural Language Processing, Algorithm Analysis, Reinforcement Learning, Unsupervised Learning, High Performance Machine Learning, Data Science, Computer Systems for Data Science, Probability, Statistical Inference, Scaling LLM Systems
Tongji University
Sep 2019 – Jun 2023Core courses: Data Structures and Algorithms (C++), Machine Learning Theory, Calculus, Linear Algebra, Statistics, Discrete Math, Numerical Methods and Algorithms, Software Engineering, Foundation of Database, Micro-service and Web Service, Data Visualization
Publications
Position: Multimodal LLMs Should Learn from Children, Yifan Shen, Chuanmiao Dong, Zhuoqing Zhong, Pei Tian, Tianjiao Yu, Jiateng Liu, Bowen Fang, Xinzhuo Li, Yuanzhe Liu, Zhengyuan Li, Chen Fang, Xu Cao, CVPR Workshops, 2026.
CogniRoute: Learning to Route Social Evidence in Omni-Modal Models, Yifan Shen, Pei Tian, Xinzhuo Li, Bowen Fang, Shujun Xia, Bingxuan Li, Ana Jojic, Wenming Ye, Xu Cao, James Matthew Rehg, Ismini Lourentzou, NeurIPS, 2026.
RIMS: Preference Optimization via Smoothed Multi-pair Aggregation for Small-Scale LLM Retrieval-Augmented Generation, Pei Tian, Zihan Dong, Tianci Liu, Linjun Zhang, Haoyu Wang, COLM, 2026.
Distributionally Robust Posterior Sampling: A Variational Approach, Bohan Wu, Pei Tian, Yixin Wang, David Blei, Journal of Machine Learning Research, 2026.
Research Experience
PLAN Lab, University of Illinois Urbana-Champaign
Dec 2025 – Present- Outperformed a proprietary model by 15.33% on social cognitive reasoning via fine-tuning a Qwen3-Omni-based MoE reasoning model.
- Improved accuracy by 10% with schema-aligned predictive routing in SFT and boosted performance by 9% with route-aware reinforcement learning.
- Curated an omni-model social cognition benchmark with grounded reasoning chains, schema labels, and temporal evidence spans.
Natural Language Text Processing Lab, Columbia University
Sep 2025 – Dec 2025- Developed a method to disentangle VLM latent space through attention-map analysis and feature decomposition with PyTorch and Hugging Face.
- Prototyped lower-level visual features in LLaVA and benchmarked visual transformer/MLP layers for artistic style analysis.
Blei’s Lab, Columbia University
Sep 2025 – Present- Designed Distributionally Robust Variational Bayes as minimax posterior inference under adversarial data-distribution perturbations across regression, probabilistic PCA, VAE, and DDPM.
- Developed Wasserstein-gradient-flow, Sinkhorn optimal-transport, and Frank-Wolfe algorithms with asymptotic and convergence guarantees.
- Demonstrated gains against nine baselines under five contamination distributions in models with and without latent representations.
Department of Computer Science, UAlbany
Apr 2025 – Present- Fine-tuned Qwen2.5 and Gemma2 with preference learning and LoRA, improving multi-hop QA by 8% accuracy and 9% F1.
- Customized DPO, ORPO, and CPO trainers; distributed training with DeepSpeed on Slurm/A100 infrastructure.
- Combined RAG, Qwen3, chain-of-thought reasoning, and rejective sampling to improve training-data quality and reliability.
DitecT Laboratory, Columbia University
Sep 2024 – Dec 2024- Developed a language-to-world pipeline coupling LLM scene generation with a diffusion video-to-video model for realistic driving scenarios.
- Applied few-shot coding agents to translate semantic constraints into executable programs across 90 scenes using OpenAI o4-mini.
- Designed multimodal conditioning for a DiT video-transfer model to preserve scene geometry and transform appearance.
Work Experience
Apple Inc.
Feb 2026 – Present- Improved Recall@10 by 15% on an agentic search pipeline for hardware development using self-evolving optimization.
- Designed a Python-native agent harness with ReAct/CodeAct controllers, unified tools, planning, context, MCP orchestration, memory, and trajectory monitoring.
- Architected heterogeneous multi-agent collaboration and a unified subagent orchestrator/runtime.
- Productionized an asynchronous FastAPI, Redis, Docker Compose, and Kubernetes agent service, achieving 1.6× acceleration through concurrent calls and multi-process execution.
Kaliber AI
Jun 2025 – Aug 2025- Achieved 99% speech-recognition accuracy and 1.3× acceleration using Triton Inference Server on Kubernetes.
- Developed VLM-enhanced ViT 3D object detection and an interactive diffusion-based conversational speech system.
- Designed a Reflexion-pattern agent system for multimodal interpretation, memory integration, and adaptive reasoning.
L’Oréal
Jan 2025 – May 2025- Led a four-person team building a knowledge-graph RAG pipeline with GPT-4o, LangChain, LlamaIndex, and Neo4j.
- Designed KNN and graph-community queries; attained 0.91 answer relevance and 0.58 faithfulness in GraphRAG evaluation.
Projects
PageAttention-Oriented LLM Inference Acceleration
Jun 2026 – Jul 2026- Implemented continuous batching with coordinated prefill/decode and a C++ Page-Attention-style KV-cache block manager.
- Enabled Qwen3 KV-cache support and cross-layer attention cache management for concurrent API requests.
Incentivizing Efficient CoT via Reward Modeling in Reinforcement Learning
Feb 2026 – Present- Fine-tuned Qwen3-1.7B with GRPO/DAPO to shorten reasoning traces while preserving accuracy using slime.
- Built SGLang-accelerated rubric judging and CodeAct sandbox execution for compact, high-quality reasoning evaluation.
Efficient Knowledge Distillation across Transformer Encoder / Decoder
Mar 2025 – May 2025- Improved small-model accuracy/BLEU with KL-based BERT/Qwen2.5 distillation and accelerated execution with Flash Attention, mixed precision, PyTorch Dynamo, and quantization.
End-to-End Machine Learning Systems Optimization from Scratch
Feb 2025 – Apr 2025- Benchmarked CUDA shared-memory tiling, coalesced access, and loop unrolling against cuDNN and Triton kernels.
- Optimized ResNet18 training and delivered 4× FP32-to-INT8 compression while preserving accuracy.
Controlling Generative Diffusion Models with Unsupervised Machine Learning
Sep 2024 – Dec 2024- Led three teammates applying PCA, ICA, MDS, random projection, and t-SNE to extract six semantic dimensions from diffusion latent representations.
- Reduced runtime by 56% using a DDIM scheduler and low-dimensional U-Net latent-space analysis.
Teaching Assistant Experience
Skills
Programming: Python, Java, C/C++, CUDA, Bash, HTML/CSS/JavaScript, TypeScript, SQL, R, Git/GitHub, APIs
Machine & deep learning: scikit-learn, PyTorch, TensorFlow, Hugging Face, Accelerate, Megatron, Ray, Lightning, DeepSpeed, Optuna, PEFT, distributed training, HPC
NLP & agents: Transformers, RNN, LSTM, BERT, GPT, T5, vLLM, SGLang, LangChain, LlamaIndex, agentic harnesses
Computer vision: Diffusion, torchvision, OpenCV, Pillow, CNN, ResNet, YOLO, U-Net, DDPM/DDIM, ControlNet, ViT, CLIP, VLM
Reinforcement learning: Policy gradient, RLHF, RLVR, PPO, GRPO, DAPO, GSPO, DPO, TRL, veRL, ms-swift, slime, agentic RL
Data & infrastructure: NumPy, pandas, PySpark, Hadoop, Docker, Kubernetes, RabbitMQ, Kafka, Flink, Neo4j, AWS, GCP, Azure, distributed systems
Honors
First Class Scholarship, Tongji University · Second Class Award, Mathematics Competition at Tongji University
MIT xPRO Certificate for Machine Learning, Modeling, and Simulation Principles