CV . Xingyue Huang

Professional Experience

Meta Platforms Inc.
Research Scientist Intern (06/2026 – 09/2026)

Developed an LLM-based survey imputation system for advertiser campaign measurement, modeling responses from 5k advertisers across 18 questions and 10 weekly waves.
Reduced aggregate distributional error by 3.4×, from 26.3% to 7.8%, using quantile-mapped KNN on Llama 4.
Designed LLM-XGBoost ensemble methods achieving the best distributional fidelity, reducing PPE to 6.7% and improving accuracy by 1.7% by using LLM predictions as tabular features.

AITHYRA Research Institute for Biomedical Artificial Intelligence
Predoctoral Fellow (03/2026 – 05/2026)

Developed RelAgent, an LLM-based autonomous data scientist that searches over SQL feature programs and model choices for relational learning.
Achieved rank 1.00 on RelBenchV2 and 4DBInfer classification benchmarks, improving average AUROC over KumoRFM-v2 from 85.91 to 87.32 and from 79.96 to 81.38, respectively.
Designed deterministic inference without further LLM calls, yielding interpretable SQL-defined feature maps paired with classical predictors for scalable database deployment.

Snap Inc.
Research Intern (06/2025 – 10/2025)

Developed Threshold Differential Attention (TDA), a sink-free, ultra-sparse attention mechanism for long-context LLMs, achieving >99% exact zeros while matching Softmax accuracy on QA benchmarks.
Pre-trained large language models from scratch and demonstrated long-context robustness on SCROLLS passkey retrieval, where TDA outperformed Softmax by approximately 2.5× at 4k-token contexts.
Co-authored Hierarchical Token Prepending, improving long-document embeddings in decoder-only LLMs via block-level summary tokens, with 5% gains across 11 retrieval datasets and 30 embedding benchmarks.

Eigent-AI
Research Intern (10/2024 – 06/2025)

Built a tool-use synthetic data generation pipeline producing 20k verified execution traces for CAMEL-AI.
Used back-translated tool trajectories for supervised fine-tuning, improving math benchmark accuracy by 5%.
Led the Loong verifier-driven RL framework for long chain-of-thought synthesis.

Service

Reviewer: NeurIPS 2025 · ICML 2026 · ICLR 2026

Lead Organizer: Workshop on Graph Foundation Models, ICML 2026

Organizer: Scaling Environment of Agents (SEA) Workshop, NeurIPS 2025

Invited Talks:

RelAgent: LLM Agents as Data Scientists for Relational Learning — Stanford University; CAMEL-AI
Graph Foundation Models Tutorial — Learning on Graphs 2025
How Expressive are Knowledge Graph Foundation Models? — Snap Inc., 2025
Relational Hypergraphs for Knowledge Graph Foundation Models — TU Wien, 2026

CV