@vamvas.bsky.social
Natural Language Processing (#NLProc #NLP) Researcher @ University of Zurich @cl-uzh.bsky.social vamvas.ch
💥Introducing new paper: arxiv.org/pdf/2510.17715, QueST — train specialized generators to create challenging coding problems.
From Qwen3-8B-Base
✅ 100K synthetic problems: better than Qwen3-8B
✅ Combining with human written problems: matches DeepSeek-R1-671B
🧵(1/5)
Sina Ahmadi receiving award.
Congratulations to @sinaahmadi.bsky.social and co-authors for receiving an ACL 2025 Outstanding Paper Award for PARME: Parallel Corpora for Low-Resourced Middle Eastern Languages!
aclanthology.org/2025.acl-lon...
Sina Ahmadi, Rico Sennrich, Michelle Wastl, Jannis Vamvas, Andrianos Michail from the Department of Computational Linguistics posing in front of the at ACL 2025 Vienna step-and-repeat banner
We're proud to see our researchers sharing their latest work: @sinaahmadi.bsky.social, @ricosennrich.bsky.social, @michellewastl.bsky.social, @vamvas.bsky.social and Andrianos Michail (pictured) and many others.
If you’re at #ACL, stop by to learn more!
Arrived in Vienna ready for #ACL2025NLP!
Excited to present papers with @vamvas.bsky.social @ricosennrich.bsky.social on Unsupervised Translation Direction Detection and Multilingual Hallucination Detection!
Come say hi! 👋
#NLProc #NLP #NMT #LLMs
Make a guess: How does GPT-4o tokenize the Lithuanian word ‘nebeprisikiškiakopūsteliautum’? A. ⟨nebepr, is, ik, iškiak, opūste, liautum⟩ B. ⟨neb, pre, ski, kayak, opus, tell, autumn⟩ C. ⟨ne, be, pris, iki, ški, ak, op, ū, stel, ia, utum⟩
Can you solve my exam question?
I've come to believe that multiple-choice exams are underrated. More in my blog post, “The Joy of Multiple-Choice.” vamvas.ch/the-joy-of-m...
I had the pleasure of presenting our paper “20min-XD: A Comparable Corpus of Swiss News Articles (arxiv.org/abs/2504.21677) ”, a joint work with @vamvas.bsky.social , Selena Calleri and @ricosennrich.bsky.social , at SwissText 2025 last week.
We are honored to receive Best Paper Award for it! ✨
Michelle Wastl at the SwissNLP Expo
Our Michelle Wastl is demonstrating her invention, the Translation Direction Detector, at the SwissNLP Expo!
Michelle's paper: arxiv.org/abs/2401.06769
Demo: huggingface.co/spaces/Zuric...
If you're at the expo, make sure to stop by the Department of Computational Linguistics UZH!
Bar chart illustrating counts of different API responses for the same request.
I also found what others complained about previously: The “Predicted Outputs” feature is unstable even with temp=0.
It’s unclear to me why this happens – do they batch the verification steps of users? An artifact of temperature sampling?
So, any hints would be appreciated! 4/4
3D plot comparing API behavior to simulation results.
Systematic probing of the OpenAI API reveals that they are using a lookahead parameter of K=16, and perform prefix matching with a prefix threshold of 32 tokens.
A simulation experiment confirms that the reverse-engineered algorithm matches the behavior of the API. 3/4
Schematic illustration of draft verification
Draft verification is a well-known idea in NLP, from Speculative Decoding to algorithms leveraging a static draft, such as Aggressive Decoding and Prompt Lookup Decoding.
Clearly, “Predicted Outputs” does something similar. But how does it work, specifically? 2/4
OpenAI introduced the “Predicted Outputs” feature last November, but their documentation raises more questions than it answers. So let’s reverse-engineer the feature! 🧵1/4
platform.openai.com/docs/guides/...
Blogged: Reverse-engineering OpenAI's Speculative Decoding.
I was curious how GPT-4o can make use of predicted outputs to speed up text generation.
vamvas.ch/openai-predi...
Congratulations to Dr. @shaitarova.bsky.social , who just successfully defended her thesis on "Generated Text as the Locus of Language Contact: Exploring the Impact of Generative AI on English and German". Special thanks to the external examiner Natalia Resende!
11.04.2025 09:30 — 👍 12 🔁 3 💬 2 📌 0
I'm thrilled to share my first PhD project, a joint work with
@vamvas.bsky.social and @ricosennrich.bsky.social
Paper link:
arxiv.org/pdf/2503.10494
Long context LLMs have paved the way for document translation, but is simply inputting the whole content the optimal way?
Here's the thread 🧵 [1/n]
Tannon Kew presenting during his viva.
Congratulations to Dr. @tannonk.bsky.social, who just successfully defended his thesis on "Leveraging Data, Decoding, and Context for Controlling Text Generation from Pretrained Language Models". Special thanks to the external examiner @feralvam.bsky.social!
06.12.2024 10:34 — 👍 18 🔁 5 💬 0 📌 0
Rico Sennrich (University of Zurich) is hiring a PhD student to work on representations of knowledge in multilingual language models, starting September 2024.
Rico is my former advisor and I can greatly recommend working with him. Apply by January 4: jobs.uzh.ch/offene-stell...
Minimum Bayes Risk decoding has recently gained momentum in #NLProc.
However, a generic implementation for @huggingface.bsky.social Transformers has been missing. Check out our new 𝗺𝗯𝗿🔥 repo, which is designed to work with any model and metric on the Hub: github.com/ZurichNLP/mbr
Can NLP be used for highlighting differences between two documents?
Our EMNLP paper investigates the task of Recognizing Semantic Differences (RSD) with simple, unsupervised approaches.
• Demo: huggingface.co/spaces/Zuric...
• Paper: huggingface.co/papers/2305....
#NLProc
