#NLProc: Stop annotating randomly-sampled data! Select-then-annotate to make language models better few-shot in-context learners (ICL)! 12% gain with GPT-3 etc over 10 tasks. 10x annotation efficiency. Different from the finetuning (FT) paradigm. arxiv.org/abs/2209.01975 1/5

My PhD thesis, "Transfer Learning in NLP through Interactive Feedback" is now online drum.lib.umd.edu/handle/1903/29… Grateful for the support from Jordan, my committee, and @ClipUMD. PhD has truly been an unforgettable journey! Now I am in NY at AWS as Applied Scientist 🗽

📢 A 🧵 on the Trends in NLP Datasets. What’s changed since SQuAD was all the rage in 2016? A: A LOT. 🔭 1. Generic ➡️ Niche Tasks 2. Task-specific Training+Eval ➡️ Eval Only 3. Dataset ➡️ Benchmark ➡️ Massive Collections 4. Datasets ➡️ Diagnostics 1/

Hokkien is the second most popular dialect in Taiwan, one of the most linguistically diverse places in the world. Some say that Taiwan is the “Urheimat” of the Austronesian languages family. Would love to see AI developed to help preserve these beautiful, spoken languages.

Our researchers won a Best Paper Award AACL 2025 for their work to make visual question answering (VQA) systems more effective for blind users. The paper was coauthored by Yang (Trista) Cao, Kyle Seelman, Kyungjun Lee and Hal Daumé III. Learn more: go.umd.edu/cDe

These are the most impressive chats we've seen with ChatGPT so far. It can...

Really exciting to see transformers being used for earth science!

Toolformer: Language Models Can Teach Themselves to Use Tools Presents Toolformer, a model trained to decide which APIs to call, when to call them, what arguments to pass, and how to best incorporate the results into future token prediction. abs: arxiv.org/abs/2302.04761

🚨BREAKING NEWS🚨 Looking to work on LLM research this summer? Our team in Amazon Web Services AI Research and Education has reopened intern hiring and we're looking for talented PhD interns! DM me your resume if you work on robustness or any of the topics below 👇🏻

Debate of the century: Chomsky vs. ChatGPT

We’re announcing Amazon Bedrock—giving customers the easiest way to build and scale generative AI applications with access to the leading foundation models including Anthropic, AI21 Labs, Stability AI and Amazon Titan FMs. Choice rules! aws.amazon.com/blogs/machine-…

*Human feedback* was the necessary secret sauce in making #chatgpt so human-like But what exactly is feedback? And how can we leverage it to improve our models? Check out our new survey on the use of (human) feedback in Natural Language Generation! arxiv.org/abs/2305.00955 1/16

Check out this large, multilingual dataset with passages from millions of Wikipedia articles and material from their cited sources!

Do multilingual language models think better in English? 🤔 Yes, they do! We show that using an LLM to translate its input into English and performing the task over the translated input works better than using the original non-English input! 😯 arxiv.org/abs/2308.01223

Do language models have an internal world model? A sense of time? At multiple spatiotemporal scales? In a new paper with Max Tegmark we provide evidence that they do by finding a literal map of the world inside the activations of Llama-2!

Vector databases have raised billions offering hosted management of text embeddings. What are you revealing in these vectors ? 🔒 We demonstrate a practical method for full recovery of 90% of sentence length embeddings (arxiv.org/abs/2310.06816).

This new paper with Santosh Vempala gives a simple statistical justification for why and when Language Models *should* hallucinate using standard pretraining, even under ideal in-distribution training conditions. [1/7] arxiv.org/abs/2311.14648

Do the best translations go beyond literal meaning? Excited to share our work at #EMNLP2023 on Automatic Explicitation in translation w Marine Carpuat and Jordan Boyd-Graber! Check out our poster session on Dec 8th Fri (today!) at East Foyer from 2pm to 3:30pm. In our paper, ...

Stories have 6 primary arcs: • “Rags to riches” (rise) • “Tragedy” (fall) • “Man in a hole” (fall-rise) • “Icarus” (rise-fall) • “Cinderella” (rise-fall-rise) • “Oedipus” (fall-rise-fall) Programmatic analysis validates Kurt Vonnegut's legendary rejected masters thesis.

Automating AI research is exciting! But can LLMs actually produce novel, expert-level research ideas? After a year-long study, we obtained the first statistically significant conclusion: LLM-generated ideas are more novel than ideas written by expert human researchers.