LLM systems · Post-training · AI for science

Jiayu Qin 秦嘉雨

Ph.D. student building faster, more reliable large-model systems.

I work at the intersection of efficient LLM serving and learning—multi-agent inference, KV-cache optimization, and multi-reward post-training—while developing foundation-model methods for scientific and financial applications.

Currently an Applied Scientist Intern at Amazon and a CSE Ph.D. student at the University at Buffalo.

Email me View CV
Google Scholar GitHub X
Jiayu Qin standing by the shoreline at sunset
3.9×
RebaseKV · LLM Serving Systems TTFT reduction with optimistic cache alignment
+5.21
BETA-RL · LLM Post-Training & RL average Pass@1 with reward conditioning
40+ years
FinTOA · Financial AI & Market Prediction of financial news transformed into market signals

About

Systems constraints meet learning objectives.

My research connects model quality with the realities of deployment: latency, memory bandwidth, cache reuse, reward uncertainty, and failure modes that only appear in end-to-end systems.

In LLM serving, I study how to reuse and compress KV states without sacrificing role fidelity or answer quality. In post-training, I design principled multi-reward methods for GRPO/GDPO-style optimization. My broader work applies representation learning, optimal transport, and multimodal modeling to molecules, proteins, and financial markets.

LLM serving systems Post-training & RL KV-cache optimization Scientific ML Representation learning

Updates

Recent news

  1. Joined Amazon as an Applied Scientist Intern, working on efficient LLM and LLM-agent inference.

  2. GeoMamba-SE(3) accepted to AISTATS 2026, and KGOT accepted to ICLR 2026.

  3. RebaseKV and BETA-RL are under review.

  4. FinTOA received USD 100K in support from Empire AI.

  5. Probability Contrastive Learning accepted to NeurIPS 2024.

Selected research

Current work

Three projects that define my current research direction across systems, learning, and applications.

LLM servingUnder review · 2026

RebaseKV: Optimistic Cache Alignment for Multi-Agent LLM Serving

A training-free architecture that decodes from inherited, semantically relevant KV states while the exact prompt-conditioned cache is materialized asynchronously, then rebases onto the fresh cache.

  • Up to 3.9× TTFT reduction across GSM8K, MMLU, and HumanEval.
  • Preserves answer-level accuracy and restores role-specific programming behavior from 5% to 97%.
Multi-agent servingKV-cache reuseInference latency

Manuscript under review

LLM post-trainingUnder review · 2026

BETA-RL: Bayesian Reward Weighting for Multi-Reward LLM RL

An uncertainty-aware framework that learns simplex-constrained precision weights for multiple rewards in GRPO/GDPO-style post-training, with an identifiability fix and entropy regularization.

GRPO / GDPOReward modelingBayesian learning

Manuscript under review

Abstract visualization of financial narratives and market signals Financial topic attention
1.25Sharpe ratio
Financial AI2026

FinTOA: LLM Agent for Financial News Topic Extraction and Market Prediction

An LLM-driven topic-attention framework that converts more than four decades of financial news into interpretable time-series signals for historical return prediction.

LLM agentsTime seriesAsset pricing
Read paper

Publications

Selected papers

Full list on Google Scholar
Diagram of molecular contrastive learning and 3D rotations AISTATS 2026

Molecular representation learning

From Transformers to State Spaces: GeoMamba-SE(3) for Fast and Accurate Molecular Learning

Jiayu Qin, Zhengquan Luo, Jian Chen, Xuhui Li, Jiayi Chen, Zhiqiang Xu

A linear-time Mamba architecture with local-frame scalarization, dual-stream selective state-space modeling, and statistically enforced SE(3) invariance.

Paper Code
Abstract molecule-protein interaction network ICLR 2026

Biomedical foundation models

KGOT: Unified Knowledge Graph and Optimal Transport Pseudo-Labeling for Molecule–Protein Interaction Prediction

Jiayu Qin, Zhengquan Luo, Guy Tadmor, Changyou Chen, David Zeevi, Zhiqiang Xu

A multimodal framework that integrates biomedical knowledge graphs with optimal-transport pseudo-labeling to improve interaction prediction and zero-shot generalization.

Paper Repository
NeurIPS 2025 Workshop

Efficient LLM inference

Towards Dynamic KV-Cache Compression: Fine-Grained Evaluation of Key and Value Ranks in LLMs

Jian Chen, Zhuoran Wang, Jiayu Qin, Ming Li, Meng Wang, Changyou Chen, Yin Chen, Qizhen Weng, Yirui Liu

An incremental SVD framework and normalized effective-rank analysis for measuring data-dependent, layer-wise KV-cache compressibility under realistic serving workloads.

Paper
Abstract graph and probability distribution visualization NeurIPS 2024

Graph contrastive learning

A Probability Contrastive Learning Framework for 3D Molecular Representation Learning

Jiayu Qin, Jian Chen, Rohan Sharma, Jingchen Sun, Changyou Chen

A Bayesian weighting framework that identifies and down-weights false positive and negative pairs caused by stochastic molecular augmentations.

Paper

Background

Experience & education

Research experience across large-model systems, scientific ML, and production machine learning.

01

Experience

  1. A

    Applied Scientist Intern

    Amazon

    Efficient LLM and LLM-agent inference, including KV-cache analysis, compression, and deployment-style evaluation.

  2. H

    Visiting Researcher

    Harvard University

    Efficiency and reliability of large foundation models in realistic drug-discovery serving environments.

  3. M

    Visiting Researcher

    MBZUAI

    Mamba- and Transformer-based molecular representation learning with memory- and throughput-aware sequence modeling.

  4. G

    Machine Learning Engineer Intern

    Galasports

    Prediction models and production pipelines for large-scale user-behavior streams, with an emphasis on monitoring and latency.

02

Education

  1. UB

    Ph.D. in Computer Science & Engineering

    University at Buffalo, SUNY

    Research in efficient large-model systems, post-training, and scientific machine learning.

  2. F

    B.Eng. in Electrical Engineering

    Fudan University

    Second-class scholarship, top 5%.

Technical toolkit

LLM systems

vLLM, SGLang, KV-cache internals, prefix reuse, LoRA, quantization

Post-training

GRPO, GDPO, reward modeling, multi-reward RL, Hugging Face TRL

Systems

Python, PyTorch, CUDA, Triton, DeepSpeed, Hugging Face

ML & theory

Representation learning, graph learning, optimal transport, multimodal ML

Contact

Let’s talk about efficient LLM systems.

I am always interested in research conversations around serving, post-training, and AI for scientific applications.

jiayuqin@buffalo.edu Download CV