Tianhao Gao

Personal Homepage

Tianhao Gao

I am an algorithm engineer at JD.com, where I build JD's AI Product Graph and item-understanding systems. My work focuses on large-scale LLM/VLM training and inference, reinforcement learning methods, representation learning, retrieval, and industrial NLP systems for e-commerce.

  • Algorithm Engineer, JD.com, 2022-Present
  • M.S. in Computer Technology, Peking University, 2019-2022
  • B.S. in Software Engineering, Wuhan University of Technology, 2015-2019

gaotianhao@pku.edu.cn | GitHub | Google Scholar

News

Publications

Conference Papers

ICML 2026

InstEmb: Instruction-Following Embeddings through Glimpses of the Future

Tianhao Gao, Jun Fang, Xiaohui Zhang, Zhiyuan Liu, Chao Liu, Pengzhang Liu, Qixia Jiang

Instruction-following embeddings with output-aware semantic signals and dual-anchor pooling.

ICML Page Project Page

InstEmb framework
AAAI 2026

FANoise: Feature-Adaptive Noise Driven Representation Learning

Jiaoyang Li, Jun Fang, Tianhao Gao, et al.

Feature-adaptive noise for robust representation learning.

Paper

FANoise overview
COLING 2022

LEGO-ABSA: A Prompt-based Task Assemblable Unified Generative Framework for Multi-task Aspect-based Sentiment Analysis

Tianhao Gao, Jun Fang, Hanyu Liu, et al.

A unified prompt-based generative framework for multiple aspect-based sentiment analysis tasks.

Paper Slides

LEGO-ABSA overview

Preprints

Joint Event Extraction via Structural Semantic Matching

Haochen Li, Tianhao Gao*, Weiping Li, et al. (* equal contribution)

arXiv

Selected Projects

JD Oxygen AI Item Center (Oxygen AIIC) / Large Product Understanding Model

Built and iterated a large-scale LLM/VLM-centric item-understanding platform centered on JD's AI Product Graph.

  • Delivered Oxygen AIIC V1 for item understanding, management, and downstream applications.
  • Reached 90%+ accuracy in product-understanding tasks through continuous model and data iteration.
  • Accumulated large-scale data assets across taxonomy, attributes, knowledge graphs, and titles, adding over 1TB of structured assets.

JD Front-end Category Recognition

Upgraded category-recognition models and data assets for large-scale product taxonomy systems.

  • Fine-tuned Qwen2.5-7B for apparel category recognition with about 90.5% validation accuracy.
  • Reduced dependence on manual labeling through automatic hard-example mining.
  • Mined product terms from large-scale search query logs for downstream taxonomy enrichment.

Product Detail Page QA / RAG

Developed retrieval-augmented QA models for product-detail scenarios.

  • Constructed SFT and evaluation data with GPT-assisted pipelines.
  • Built a LangChain-based QA flow and deployed it for multiple first-level categories.
  • Improved vector retrieval top-1 accuracy from 76.24% to 77.29%.

Apple Store Customer-Service RAG

Built and evaluated RAG components for Apple self-operated store customer-service scenarios.

  • Reached 85%+ offline accuracy and 89.5% including partially correct UAT answers.
  • Validated Qwen-series baselines and prepared incremental model upgrades.
  • Designed an LLM-based information-extraction pipeline from customer-service conversations.

Counterfeit Brand Title Detection

Designed a modular detection system for counterfeit-brand title risks.

  • Decoupled detection challenges into explainable modules.
  • Combined edit distance, minimum-window matching, recursive algorithms, and CV2-based signals.

Multimodal Product Content Tagging

Built a multimodal tagging pipeline for e-commerce product understanding.

  • Compared base models and architectures for production constraints.
  • Constructed high-quality training data with PySpark and Hive SQL.
  • Reached about 90% accuracy with a recursive thresholding strategy.
Multimodal product tagging

Patents