I am a senior research scientist at Google DeepMind in New York, working on natural language processing.
For most of the last decade my research centered on paraphrase and representations of meaning, including the earliest work on general-purpose sentence embeddings trained for semantic similarity — along with the evaluations used to measure them. That thread branched widely: synthetic paraphrase data at scale (ParaNMT-50M), controllable generation and adversarial examples, style transfer, and — to my knowledge, a first — reinforcement learning with neural metrics as the reward for machine translation. A recurring theme is simple methods that stay ahead for years. During those years I also advised QuillBot on its NLP technology, from shortly after its founding through to its acquisition.
At Google I work on pre-training data quality — scoring and filtering the data behind Gemini, including detecting and removing AI-generated content — and, most recently, reinforcement learning and self-training for sign language translation, returning to the RL-for-translation ideas I first explored in 2019. That work shipped as SL2T, the model that now powers sign-to-text on Pixel and the first sign language AI in a consumer product. Models I built to detect AI-generated content have launched in Google Search to demote AI-generated websites, and in YouTube and Google Reviews to remove AI-generated comments and reviews. I have also studied how paraphrasing evades AI-text detectors (NeurIPS 2023), and worked on watermarking language models, automatic attribution, and long-form generation, contributing along the way to the Gemini, Gemma, and PaLM model families.
I did my Ph.D. at Carnegie Mellon, advised by Taylor Berg-Kirkpatrick and Graham Neubig, working closely with Kevin Gimpel at TTIC. Before that I did an MS with Dan Roth at Illinois, and undergraduate degrees in mathematics and chemistry at Wisconsin.
The best way to reach me is by email.
News
- New preprint: POLARIS, on guiding small models to write long stories.
- StoryScope, investigating idiosyncrasies in AI fiction, accepted to COLM 2026.
- Two watermarking papers accepted to TMLR.
Publications
-
51 · 2026
POLARIS: Guiding Small Models to Write Long Stories
-
50 · 2026
StoryScope: Investigating Idiosyncrasies in AI Fiction
-
49 · 2026
Improving Detection of Watermarked Language Models
Tackles the detection side of watermarking, showing that combining watermark-based detection with model-based detectors outperforms either alone — including on text from unwatermarked models.
-
48 · 2026
A Watermark for Black-Box Language Models
A watermarking scheme requiring only sampling access to a model — no logits — making watermarking practical for API-only settings while retaining distortion-free guarantees.
-
47 · 2026
Gemma 4 Technical Report
-
46 · 2025
A Gold Standard Dataset for the Reviewer Assignment Problem
-
45 · 2025
Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
-
44 · 2025
Gemma 3 Technical Report
-
43 · 2024
PostMark: A Robust Blackbox Watermark for Large Language Models
- 42 · 2024
-
41 · 2024
Multiple References with Meaningful Variations Improve Literary Machine Translation
-
40 · 2024
Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context
- 39 · 2023
-
38 · 2023
Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense
Showed that a discourse-level paraphraser (DIPPER) evades watermarking, GPTZero, DetectGPT, and OpenAI's own classifier — and that retrieval over previously generated text is the one defense that holds up. Widely cited in the debate over whether AI-text detection is possible at all.
- 37 · 2023
- 36 · 2023
- 35 · 2023
-
34 · 2023
FIAT: Fusing Learning Paradigms with Instruction-Accelerated Tuning
Fused the two dominant learning paradigms — in-context instruction and parameter tuning — into a single method that outperforms both ICL and fine-tuning alone across model scales.
- 33 · 2023
- 32 · 2023
-
31 · 2023
Gemini: A Family of Highly Capable Multimodal Models
-
30 · 2023
PaLM 2 Technical Report
-
29 · 2022
Canine: Pre-training an Efficient Tokenization-Free Encoder for Language Representation
The first pre-trained tokenization-free encoder: CANINE operates directly on characters, sidestepping the brittleness of fixed subword vocabularies while outperforming a comparable mBERT model with fewer parameters.
- 28 · 2022
- 27 · 2022
- 26 · 2022
- 25 · 2022
- 24 · 2022
- 23 · 2021
- 22 · 2021
- 21 · 2020
-
20 · 2020
Reformulating Unsupervised Style Transfer as Paraphrase Generation
Reframed unsupervised style transfer as paraphrase generation (STRAP): normalize the style away with a paraphraser, then style it back. No parallel data or style-specific engineering, with large gains over prior systems across formality, Shakespeare, and social-media styles.
- 19 · 2020
-
18 · 2020
Learning and Applications of Paraphrastic Representations for Natural Language
- 17 · 2019
- 16 · 2019
- 15 · 2019
- 14 · 2019
-
13 · 2018
ParaNMT-50M: Pushing the Limits of Paraphrastic Sentence Embeddings with Millions of Machine Translations
Showed that back-translating a bitext yields tens of millions of paraphrase pairs essentially for free, and that simple embeddings trained on them outperform far more sophisticated architectures — a baseline the Sentence-BERT paper later acknowledged as remarkably strong. The dataset became a standard resource for paraphrase generation and semantic similarity.
-
12 · 2018
Adversarial Example Generation with Syntactically Controlled Paraphrase Networks
Introduced syntactically controlled paraphrase networks, which rewrite a sentence into a specified syntactic form — and showed these adversarial rephrasings break models that word-level perturbations leave intact.
- 11 · 2018
- 10 · 2018
- 9 · 2017
- 8 · 2017
-
7 · 2016
Towards Universal Paraphrastic Sentence Embeddings
The earliest general-purpose sentence embeddings trained for semantic similarity, along with the evaluation framework — direct evaluation on semantic textual similarity benchmarks — that later work adopted wholesale, including SentEval and the sentence-embedding evaluations that followed it. Its finding that averaged word embeddings outperform LSTMs became the basis for the “tough-to-beat baseline” of Arora et al. (2017).
- 6 · 2016
- 5 · 2016
-
4 · 2015
From Paraphrase Database to Compositional Paraphrase Model and Back
The Paragram vectors trained here — including the 300d Paragram-SL999 set released in the appendix — were the first to reach human-level agreement on SimLex-999, and became a standard initialization for later similarity models.
-
3 · 2015
Clustering With Side Information: From a Probabilistic Model to a Deterministic Algorithm
-
2 · 2014
Tiered Clustering to Improve Lexical Entailment
-
1 · 2013
Illinois Cognitive Computation Group UI-CCG TAC 2013 Entity Linking and Slot Filler Validation Systems
Software & data
-
data
ParaNMT-50M
-
models
Paraphrastic representations at scale
-
tool
compare-mt
Service
Area Chair
- AAAI2022, 2023, 2026 · Senior Program Committee, 2027
- NeurIPS2023, 2024, 2025
- ICML2025
- ACL2020, 2024
- NAACL2021, 2024
- EMNLP2024
- EACL2024
- ARRongoing
Reviewing
- TMLRongoing
Other
- ACL 2020Paper Matching System — designed and trained the abstract similarity model used to assign reviewers to papers, and helped build the system with Graham Neubig
Invited talks
- Georgia Tech — CS 4650, Natural Language Understanding
- Google — NLX invited talk · Google — Lux Leads meeting
- Ohio State University · UMass Amherst · Allen Institute for AI · IBM
- Midwest Speech and Language Days, University of Chicago
- Midwest Speech and Language Days, University of Chicago · CCICADA Research Retreat, Carnegie Mellon