Xinyu Guan
2 Papers
3 Patents
5+ Yrs Exp

AI Agent Researcher · TaoTian Group @ Alibaba

Open to Collaborate

关鑫宇

AI Agent · Time-Series Forecasting · LLMs · Multilingual NLP · RLHF Alignment

I am an AI Agent Researcher at Alibaba TaoTian Group, building predictive AI Agents for logistics & supply-chain optimization. Previously at Tencent Hunyuan, where I worked on MMLU, CMMLU and other benchmark improvements, and Baidu ERNIE Core Team, where I focused on multilingual capabilities and LLM Arena rankings.

I hold an MSc in Computer Science from the University of Glasgow (USNews #62), where I worked with Prof. David Manlove (University of Oxford) on efficient text search algorithms. My research interests lie at the intersection of AI Agents, scalable LLM training, alignment techniques, and multilingual NLP.

Contact Me Publications Experience
Large Language Models
Multilingual NLP
RLHF / Alignment
Data-Centric AI
Web-Scale Parsing
Recommendation Systems

# News

Feb 2026

Joined Alibaba TaoTian Group as AI Agent Researcher, working on AI Agent applications for logistics and supply-chain optimization.

Nov 2025

ERNIE EB5 tied #2 globally on LMArena (China #1), surpassing DeepSeek – Text Arena win rate 86.4%, multilingual avg accuracy 92%.

Nov 2025

Vision-Arena: EB5-VL achieved +25.5% win rate vs baseline, tied #1 industry-wide.

Nov 2025

Patent filed: "A Multi-Agent and LLM Collaborative Multi-Dimensional Text Quality Scoring System" (First Inventor, under review).

Oct 2025

Paper "Evaluation and Optimization of Efficient Text Search Algorithms" submitted to ICASSP 2026 (First Author, under review).

Mar 2025

Paper "Price-Aware Dynamic Heterogeneous Hypergraph Network for Next Basket Recommendation" published at ICASSP 2025 (Second Author).

# Current Work

Time Agent NeurIPS 2026 Target

AI Agent for time-series forecasting, combining foundation models (TimeGPT/TimesFM-style) with LLM reasoning for accurate predictions across diverse domains.

Multi-Agent Orchestration Work Project

Designing modular agent frameworks with episodic/semantic memory and hierarchical planner-executor-reflector patterns for supply-chain optimization.

Xianyu Quality Inspection Work Project

AI Agent for photo compliance detection and physical defect inspection on second-hand products.

# Selected Publications & Patents

Conference Papers

Suffix Tree Generation
Under Review

Evaluation and Optimization of Efficient Text Search Algorithms

Xinyu Guan

ICASSP 2026 First Author · Oct 2025

Evaluated 7 classical string-matching algorithms on large corpora; proposed a suffix-tree-based algorithm combined with Ukkonen's construction, achieving ~5.2× speedup on large-scale datasets. Targeted human gene sequence search with 40% speed improvement on 3 billion sequences.

Heterogeneous Hypergraph
Published

Price-Aware Dynamic Heterogeneous Hypergraph Network for Next Basket Recommendation

First Author, Xinyu Guan (Second Author)

ICASSP 2025 Second Author · Mar 2025

Proposed a dynamic heterogeneous hypergraph network that incorporates price signals into next-basket recommendation, capturing complex item correlations and temporal purchase dynamics.

Patents

Under Review

A Multi-Agent and LLM Collaborative Multi-Dimensional Text Quality Scoring System

First Inventor · Nov 2025

Granted

A Generative Large Model Watermarking Tool Based on Probability Perturbation Encryption

First Inventor · Jul 2024

Granted

A Database Drag Behavior Detection Method Based on Time Series

Co-Inventor · Jun 2024

# Work Experience

Alibaba Group

TaoTian Group — AI Agent Researcher

Feb 2026 – Present
Xianyu Quality Inspection Building AI Agents for photo compliance detection and physical defect inspection on second-hand products.

Baidu

ERNIE Foundation Model Core Team — Research Scientist (Senior)

Oct 2025 – Dec 2025
ERNIE Alignment & Multilingual Post-Training Led DAPO-based post-training alignment for EB5 on PaddlePaddle, scaling to 260 H800 GPUs across 80+ languages; designed SOTA multilingual data scaling for hard prompts and low-resource languages. Achieved Text Arena win rate 86.4%, multilingual avg accuracy 92%; tied #2 globally on LMArena (China #1), surpassing DeepSeek.
Instruction Evolution & Data Enhancement Led instruction evolution across 10 innovation axes with history-aware learning; built a high-quality automated self-play best-of + human N-to-1 pipeline for QA/creative/multimodal. On Vision-Arena: +25.5% win rate vs baseline, tied #1 industry-wide.

Tencent

Hunyuan Text-to-Text Pipeline Team — Research Scientist

Mar 2025 – Sep 2025
Yuanbao AI Search Trained web page classification & main-text extraction models (NeuroBERT-pro + GATv2Conv) on 8×H100; deployed on 800×H100 GPUs at ~50 billion HTML pages/day. Improved extraction accuracy to 65%, classification accuracy 70%→80%, structural segmentation 60%→65%.
Maximum-Rectangle Recall Algorithm Designed a novel recall algorithm with HTML2Text formatting; boosted Precision 83%→98%, improved Recall ~1.5%.
HTML Pruning (8 iterations) Improved recall 73.41%→98.33% with 57.14% compression, reducing downstream pre-training tokens by >50%.
Multilingual Evaluation Arena-style evaluation against SOTA LLMs for Southeast Asian languages; produced ≈900B multilingual tokens via ELO-guided instruction evolution.

Tencent

Hunyuan Strategy Group 4 — Research Scientist

Feb 2024 – Mar 2025
Biomedical Domain Enhancement GPT-4o–based agent pipeline with 99% concept-extraction precision; curated 934 high-quality questions. MMLU Biomedical +11.7 pts, CMMLU +10.2 pts.
Video Processing Pipeline Integrated PySceneDetect + Video-LLaMA; reduced 10-min 1080p processing to ~80s; capacity ~20,000 video-hours/month.
Data Recognition Model Refined Magika model to >99% file-type accuracy across 116 file types; deployed at ~10 QPS (800k items/day).
Audio Synthesis & Alignment Whisper-based sliding-window alignment at 99% text-audio accuracy; delivered 20k+ high-quality transcripts (patented).

Institute of Information Engineering, Chinese Academy of Sciences

Research Assistant Intern

Nov 2023 – Jan 2024
Insider Threat Detection (LLM4ITD) Fine-tuned OPT-125M, ChatGLM-6B (P-tuning), LLaMA-7B (LoRA) on CERT 4.2; achieved AUC ≈ 0.99, Recall ≈ 0.99, FPR ~0.3%. LoRA/P-tuning reduced training time ~30% and compute >50%.
Knowledge Graph Construction Curated 20,000+ RDF triples; 98% extraction accuracy, 43% higher precision than rule-based baseline.

Mico World

Yoho Department — Software Engineer

May 2021 – May 2022
H5 Architecture Led H5 architecture for Greedy Lion (~20% of company revenue); critical path optimization improved page load speed 3.7×.
WakaDashboard (Vue.js) Designed enterprise management system from scratch; delivered >80% of core APIs; boosted frontend business efficiency 73%.

# Research Experience

LLM4ITD: Insider Threat Detection with Fine-Tuned LLMs

Nov 2023 – Feb 2024

Reconstructed classification tasks into QA framework across multiple 7B models. Structured prompt templates for complex user behavior recognition. AUC 0.9923, Recall 98.83% on CERT dataset. LoRA/P-tuning on ~0.01% parameters: training time −30%, compute −50%.

Efficient Text Search Algorithm Evaluation

Jun – Oct 2023

Collaborated with Prof. David Manlove (University of Oxford). Evaluated 7 classical algorithms (KMP, BM, suffix tree, Ukkonen, etc.) on 1M+ samples from NLTK corpora. Proposed suffix-tree + Ukkonen approach: ~5.2× speedup. Gene sequence search: 40% faster on 3B sequences, 5× on 20M test set at 100% accuracy.

EEG Feature Analysis for SCI Patients (KNN / SVM)

Jan – Jun 2023

Classified central neuropathic pain in spinal cord injury patients. Integrated variance thresholding, chi-square, RFE, and local linear embedding with KNN and SVM on 5,000 EEG samples. KNN accuracy 69.4%→77.8% (+8.4 pts); SVM accuracy 94.4% with mitigated overfitting.

ML Analysis of WSI Colorectal Cancer Datasets

Nov 2022 – Jan 2023

Applied BERT-based data augmentation (~25% of dataset) to whole-slide imaging. Evaluated K-Means, Louvain, PCA, and UMAP approaches. Contributed to a related conference paper draft.

Yellow Crane Tower Tourism Dialogue System (Llama2 7B-Chat)

Aug – Oct 2023

Enhanced Llama2 7B-Chat for a tourism-domain dialogue system with fine-tuning, retrieval augmentation, and domain-specific knowledge injection.

# Education

University of Glasgow

MSc in Computer Science

Sep 2022 – Dec 2023 GPA 3.67 / 4.0 USNews Rank #62

Research collaboration with Prof. David Manlove (University of Oxford) on efficient text search algorithms.

Hubei University

BEng in Software Engineering

Sep 2016 – Jun 2020