Accepted
Nominate-then-Adjudicate has been accepted as a short paper at IEEE BigData 2026. The preprint is now available on arXiv.
Xinyu Guan 关鑫宇
I am an AI Agent Researcher at TaoTian Group @ Alibaba, where I develop general-purpose AutoResearch systems and multimodal quality-inspection agents for Xianyu. My interests span AutoResearch, post-training, and agentic reinforcement learning, with a focus on reliable long-horizon reasoning, hierarchical memory, and feedback-driven self-improvement. I enjoy turning research ideas into practical systems and welcome academic collaboration on autonomous research and reliable AI agents.
News & updates
11 updatesPreparing
Preparing PIVOT, a multimodal agent self-improvement framework, for CVPR.
Submitted
TIMBRE, my first-author time-series forecasting manuscript, has been submitted to ICASSP 2027. The preprint is publicly available on arXiv.
Show fewer
Submitted
Three ICLR 2027 submissions: VLA Early Exit, LoRA Attribution, and QESChunker.
Updated
Updated the ICONIP 2026 camera-ready version of Decision-Aware Memory Cards on arXiv.
Submitted
Nominate-then-Adjudicate, our LLM-assisted MIP solver-tuning manuscript, has been submitted to IEEE BigData 2026.
Accepted
Submitted
Submitted SILICA to the August 2026 ACL ARR cycle, with EACL as the preferred venue.
Submitted
Submitted Advantage Scale Calibration to AAAI 2027.
Preprint
Released CICL, a decision-aware memory framework for tool-using agents.
Career
Work experience
Feb 2026 — Present
TaoTian Group @ Alibaba
AI Agent Researcher · P6
AI Agent research spanning General AutoResearch and multimodal quality inspection for Xianyu.
Details
- General AutoResearch — Re-architected the Agent runtime with bounded loop control, stalled-branch rerouting, evidence-driven hypothesis generation, hierarchical memory, and resumable execution traces. The system autonomously performs bad-case diagnosis, candidate generation, and prompt or policy iteration. Across 19 business-task optimization runs, it completed prompt iteration, candidate validation, and migration validation without human intervention within the optimization loop. Across 4 Agent runtimes × 4 model configurations, the selected stack met target on 8 of 9 controlled image-QC tasks, with 279 migration tests passed.
- Xianyu Multimodal Quality Inspection — Built inspection workflows for viewpoint compliance detection, base photo-quality checks, and visible physical-defect detection. An offline acceptance snapshot limited to nine categories and 80 inspection checks reached 97.54% mean Macro-F1. Separately, the current capability supports 30+ product categories at approximately 30K orders per day.
Oct — Dec 2025
Baidu · ERNIE Foundation Model
Senior Research Scientist · T4+
DAPO-based post-training and alignment for EB5 on PaddlePaddle.
Details
Baidu / ERNIE Foundation Model Core Team
- Scaled multilingual post-training to 260 H800 GPUs across 80+ languages; reached a Text Arena win rate of 86.4% and multilingual average accuracy of 92%, tied global #2 and China #1 on LMArena in Nov 2025.
- Led ten-axis instruction evolution and a best-of self-play plus human N-to-1 data pipeline; achieved a +25.5% win-rate result against EB45T-VL-SFT-v13 and tied #1 on the internal Vision-Arena.
Feb 2024 — Sep 2025
Tencent · Hunyuan Foundation Model
Research Scientist · T5
Mathematical and biomedical capability enhancement, pre-training data, multilingual capability improvement, and Yuanbao AI Search.
Details
Tencent / Hunyuan Text-to-Text Pipeline Team
Mar 2025 — Sep 2025
Research Scientist · T5
Yuanbao AI Search with NeuroBERT-pro and GATv2Conv.
- Trained on 8 H100 GPUs and deployed on 800 H100 GPUs, processing approximately 50B input tokens/day; extraction reached 65%, classification improved 70% → 80%, and structural accuracy improved 60% → 65%.
- Improved maximum-rectangle precision 83% → 98% and recall +1.5%; HTML pruning raised recall 73.41% → 98.33% with 57.14% compression and reduced pretraining tokens by over 50%.
- Produced approximately 900B multilingual tokens through instruction evolution and data synthesis.
- Developed and maintained STEM main-text extraction operators, completing 135 deliveries and ranking 2nd among 35 contributors in the closed research program.
Tencent / Hunyuan Strategy Group 4
Feb 2024 — Mar 2025
Research Scientist · T5
Mathematical and biomedical capability enhancement, video processing, data recognition, and audio alignment.
- Built a GPT-4o-based agent pipeline with 99% concept-extraction precision and curated 934 high-quality questions; MMLU improved +11.7 and CMMLU improved +10.2.
- Processed a 10-minute 1080p video in about 80 seconds, with capacity of 20,000 video-hours per month.
- Magika exceeded 99% across 116 file types, running at 10 QPS for 800,000 files/day.
- Whisper alignment reached 99% and delivered 20,000+ transcripts, contributing to a patent.
Nov 2023 — Feb 2024
Chinese Academy of Sciences
Research Assistant Intern
Fine-tuned LLM approaches for insider-threat detection and structured knowledge-graph construction.
Details
Institute of Information Engineering, Chinese Academy of Sciences
- Worked with OPT-125M, ChatGLM-6B P-tuning, and LLaMA-7B LoRA on CERT 4.2; AUC and Recall were about 0.99, FPR about 0.3%, training time 30% lower, and compute over 50% lower.
- Curated 20,000+ RDF triples with 98% extraction accuracy and precision +43%.
Earlier experience
May 2021 — May 2022
Mico World / Yoho Department
Software Engineer
Frontend architecture and production systems for international social products.
Details
- Greedy Lion H5 generated about 20% of company revenue; critical-path optimization improved load speed 3.7×.
- WakaDashboard delivered over 80% of core APIs and raised engineering efficiency 73%.
Publications & Manuscripts
PIVOT: Choosing When to Refine Prompts or Acquire Evidence for Multimodal Agent Self-Improvement
Studies when multimodal agents should refine prompts or acquire additional visual evidence, with feedback-driven intervention selection and independent verification.
Diagnostics and Infrastructure for Foundation Model-Based Multi-Agent Systems: A Review of the Runtime Stack
Reviews diagnostics, infrastructure, and the runtime stack of foundation-model-based multi-agent systems. HEP Preprint · PDF
TIMBRE: Teaching Time Series Forecasters to Read, Remember, and Reconcile
Combines source-aware evidence organization, state-conditioned historical response transfer, and reliability-guided fusion for event-informed time-series forecasting. arXiv · Code
How Deep Should a VLA Think When Thinking Costs Time? Budget-Constrained RL for Early Exit
Learns when a frozen vision-language-action model should exit early or keep computing, using budget-constrained RL while the simulated environment continues to evolve.
Static Gradient Attribution Underperforms a Density-Matched Random Mask on Loss-Based Forgetting Within LoRA B-Matrix
Isolates a parameter-pool confound in LoRA masking: density-matched random B-matrix masks reduce loss-based forgetting across seven seeds, while accuracy evidence remains inconclusive. arXiv · PDF · OpenReview
QESChunker: A Single Objective Unifies Overlapping and Non-Overlapping Chunking for RAG
Builds static RAG chunk indices from question–evidence supervision, unifying overlapping and non-overlapping chunking as budget-aware utility maximization. arXiv · PDF · OpenReview
Nominate-then-Adjudicate: LLM-Assisted One-Pass and Per-Instance MIP Solver Tuning
Combines fixed-portfolio nomination with structure-aware LLM adjudication for per-instance MIP solver tuning, using one model call and zero solver calls at inference. arXiv · PDF
SILICA: Certified Counterfactual Evaluation of Identifiability in Unseen-Language Induction
Evaluates whether an unseen-language task is identifiable by constructing same-state counterfactual queries with provably different correct actions.
Advantage Scale Calibration Imbalance in Group-Relative Optimization under Low-Variance Rewards: Diagnosis and Bounded Recovery
Diagnoses calibration imbalance under low-variance rewards and introduces bounded recovery for credible small reward gaps.
Decision-Aware Memory Cards: Counterfactual-Inspired Context Selection and Compression for Tool-Using LLM Agents
Introduces CICL, a decision-aware context layer that selects evidence by its expected effect on an agent's next action.
Optimizing Text Search: A Novel Pattern Matching Algorithm Based on Ukkonen's Approach
Presents a pattern-matching approach based on Ukkonen's construction for efficient text search.
Basket-Enhanced Heterogenous Hypergraph for Price-Sensitive Next Basket Recommendation
Introduces a basket-enhanced heterogeneous hypergraph for price-sensitive next-basket recommendation. arXiv
Patents
- Under ReviewA Multi-Agent and LLM Collaborative Multi-Dimensional Text Quality Scoring System
- GrantedA Generative Large Model Watermarking Tool Based on Probability Perturbation Encryption
- GrantedA Database Drag Behavior Detection Method Based on Time Series
Current research
I study how agents can research, learn, and act reliably across three core directions, with Xianyu AI as a practical application domain.
AutoResearch
Closed-loop agents for scientific discovery.
Post-Training
Data and alignment for stronger reasoning.
Agentic RL
Reliable reinforcement learning for tool-using agents.
Current manuscripts
PIVOT · Preparing for CVPR
Agent Research Survey · Submitted
More about me
I research and build reliable AI agents that can reason, learn from feedback, and turn research ideas into systems that work in the real world.
At TaoTian Group @ Alibaba, my current work focuses on AI agent research across three core directions: AutoResearch, post-training, and agentic reinforcement learning. I apply these ideas in Xianyu AI systems and am especially interested in closed-loop research agents, stable reasoning objectives, and high-quality data systems for foundation models.
Before Alibaba, I contributed to foundation-model development at Tencent and Baidu. At Tencent, I worked on the Hunyuan Foundation Model, focusing on mathematical and biomedical capability enhancement, pre-training data, multilingual capability improvement, and Yuanbao AI Search. At Baidu, I contributed to multilingual capability enhancement for the ERNIE Bot 5 (EB5) Foundation Model.
Earlier, at the Chinese Academy of Sciences, I conducted research on knowledge graphs and LLM-based security.
I am always happy to connect and collaborate on ambitious research problems. Feel free to reach out by email or visit my Google Scholar.
Research experience
Nov 2023 — Feb 2024
LLM4ITD: Insider Threat Detection with Fine-Tuned LLMs
Reformulated classification tasks as question-answering with structured prompts and parameter-efficient tuning of about 0.01% of parameters on CERT 4.2. AUC reached 0.9923, Recall 98.83%, training time was 30% lower, and compute was over 50% lower.
Aug 2023 — Oct 2023
Yellow Crane Tower Tourism Dialogue System
Extended and reparameterized RoPE and applied LoRA to Llama 2 7B. Built approximately 15,000 GPT-4-distilled tourism dialogues, with about 89% answer accuracy in human evaluation.
Jun 2023 — Oct 2023
Efficient Text Search Algorithm Evaluation
Collaboration with Prof. David Manlove. Evaluated seven algorithms on a 1M+ NLTK corpus and developed a Ukkonen-based suffix-tree approach with approximately 5.2× speedup. Applied it to human genomic sequences, improving search speed by 40% and reaching 100% accuracy in the reported evaluation.
Apr 2023 — Jul 2023
Lung Cancer Literature Classification with BioBERT
Expanded the dataset by approximately 6,000 samples through synonym replacement and back-translation, fine-tuned BioBERT with hierarchical freezing, and reached 93% PubMed accuracy with about 15% improvement on the manual evaluation set.
Jan 2023 — Jun 2023
EEG Feature Analysis for SCI Patients
Applied feature-selection methods to 5,000 samples: KNN improved 69.4% → 77.8%, and SVM reached 94.4%.
Nov 2022 — Jan 2023
ML Analysis of WSI Colorectal Cancer Datasets
Applied BERT-based augmentation to around 25% of the dataset and evaluated KMeans, Louvain, PCA, and UMAP.
Education
Sep 2022 — Dec 2023
University of Glasgow
MSc in Computer Science
GPA 3.67 / 4.0
Research collaboration with Prof. David Manlove, University of Glasgow professor and University of Oxford graduate.
Sep 2016 — Jun 2020
Hubei University
BEng in Software Engineering