Curriculum vitae

Tong Niu · mrnt0810@gmail.com

Download PDF

AI researcher with 9+ years in large language models, retrieval, and natural language generation, published at ACL, EMNLP, NAACL, ICLR, AAAI, and TACL, with embedding and reranking models ranked #1 on MTEB and BRIGHT. Takes research into production: a unified on-device speech model at Archer Aviation, and an LLM-generated regression test suite for the Salesforce Core app that catches release regressions before customers do.

Experience

Staff AI Research Engineer, Archer Aviation

2025 – 2026
  • Spearheaded the 0-to-1 design, training, and rollout of an end-to-end on-device speech model that replaced four legacy systems with higher generation quality, earning a discretionary executive RSU award for accelerated production impact.
  • Invented a timestamp-to-token alignment algorithm that removed the need for a forced aligner.
  • Optimized the model for on-device iOS inference and completed its full production deployment within latency targets.
  • Managed and mentored two direct reports, setting technical direction and guiding daily execution.

Senior Research Scientist, Meta Superintelligence Labs

2025
  • Built an OCR-augmented vision-language model pipeline that grounds homework answers in the recognized problem text.
  • Built the evaluation harness and error taxonomy separating OCR failures from reasoning failures to target fixes.
  • Improved homework-answer accuracy over the base vision-language model across question types.

Senior Research Scientist, Salesforce AI Research

2019 – 2025
  • Test generation. Generated a large regression test suite for the Salesforce Core app with LLMs, catching regressions in new releases before customers do and saving millions of dollars in manual testing per release.
  • Reranking. Built JudgeRank, an agentic LLM reranker for reasoning-intensive retrieval that held the #1 spot on the BRIGHT benchmark for six months.
  • Embeddings. Co-developed and served the SFR-Embedding model, ranked #1 on MTEB.
  • Pretraining data. Built a Selenium-based crawler for large-scale academic corpora used in LLM pretraining, improving MMLU performance.
  • LLM safety. Developed DetoxiGen, a parameter-efficient contrastive-decoding method that steers LLM output away from toxic content.
  • Applied NLP. Built HTML-aware machine translation for localization services, RAG-based question answering for banking regulatory compliance, and a recommendation system for email subject-line generation.
  • Research. Published at ACL, EMNLP, NAACL, ICLR, and ACM TOCHI on paraphrasing, cross-lingual retrieval, question generation, and dialogue evaluation.
  • Internship. Started as a Research Intern, building an unsupervised BERT-based sentence-compression model.

Research Assistant, University of North Carolina at Chapel Hill (with Prof. Mohit Bansal)

2017 – 2019
  • Adapted AutoAugment to NLP, learning data-augmentation policies for dialogue tasks (EMNLP 2019).
  • Designed adversarial strategies exposing over-sensitive, over-stable dialogue models (CoNLL 2018).
  • Built RL-based polite dialogue generation that requires no parallel data (TACL 2018).

Education

Duke University, M.S. in Computer Science (PhD program, PhD Fellowship)

2014 – 2016

University of Virginia, B.S. in Mathematics and Computer Science, minor in Physics (completed in three years)

2011 – 2014

Publications & patents

31 papers and preprints and 9 patents — see thefull list.

Skills

Engineering. Python, PyTorch, On-device (iOS) inference optimization, Model serving, Web crawling (Selenium)

Languages. English (fluent), Chinese (native)

Research areas. Large language models, Text embeddings and reranking, Retrieval-augmented generation, LLM agents, Vision-language models, Reinforcement learning, Machine translation, Speech modeling and alignment

Awards

FY25 Team Astro Award, Atlas, Salesforce AI Research

2024

PhD Fellowship, Computer Science, Duke University

2014

GPNANO Fellowship, Nanoscience, Duke University

2015

Facebook ParlAI Research Award, Facebook AI Research

2017

Books & ventures

Founder, Jake's American English — Founded and ran an English school before college, growing it to 450 students and 2 branches.

2010

Left and Right: The Conjoined Brothers — World's first graded Chinese reader series (300-word level).

2008

Study Grammar Like a Pro — A top-selling English grammar book written in Chinese.

2013