CS PhD student @UofT https://xitter.cf/t.co/cVbiZvkTHL
Joined February 2020
- Tweets75
- Following540
- Followers130
- Likes239
Markus Frohmann retweeted
1/🧵New Research on LLM-Driven Discovery!
LLMs are increasingly used to discover new algorithms, theorems, and drugs by iteratively optimizing an objective inside a "harness."
Our finding: How you initialize a discovery harness matters a lot more than you might expect.
New preprint from my @thomsonreuters Labs internship! 📄✨
LLM rerankers, reward models and multi-doc QA systems score many candidates in one prompt, so scores depend on the order 🔀 Two scorers with equal nDCG can still make different decisions.
🧵👇 arxiv.org/abs/2608.26762
📊 OC-SFT is also more stable than order-averaged distillation on 12 base models across three families.
🎯 And a single OC-SFT permutation retains sets that overlap more than ten averaged permutations of an off-the-shelf scorer.
💡 Our takeaway: a comparison of LLM scorers should report what a threshold retains and a reader answers, not ranking quality alone.
Joint work with Mahdiyar, Elizabeth, and Navid!
💻 Code: github.com/thomsonreuters/pr…
Excited to share that I'll be joining @thomsonreuters Labs in Zug, Switzerland, as an Applied AI Scientist Intern from April to September 2026! 🏔️
A nice bridge between finishing my MSc at JKU Linz and starting my PhD later this year - let me know if you're around!
wtpsplit now supports length-constrained segmentation ✂️
min/max chunk length (chars) while preserving semantic chunks - should be great for RAG!
Example (≤30 chars):
[Landing 5pm → Beimen.]
[Let's meet at: Ximen Exit 6.]
[Then: Ningxia Night Market...]
[Late-night snack!']
I'm at #EMNLP2025 in Suzhou this year!
Looking forward to connecting with the community after a year's break and spending some time abroad。。。再見！
🤖🗣️ Double Entendre will be presented today, 18:00-19:30 at Hall 4/5 by @m_schedl! Check it out if you're at #ACL2025!
Excited to share two new papers on AI-generated music detection from my research internship at @Deezer, published in @ismir_conf #ISMIR2025 and @aclmeeting #ACL2025 Findings! 🎶🤖
The problem: most AI music detectors are impractical or unreliable in real-world settings.
Excited to share two new papers on AI-generated music detection from my research internship at @Deezer, published in @ismir_conf #ISMIR2025 and @aclmeeting #ACL2025 Findings! 🎶🤖
The problem: most AI music detectors are impractical or unreliable in real-world settings.
I view this work as an important extension of current single-modality detectors while maintaining flexibility and modularity. It's not production-ready, but it highlights key paradigms for detection:
Using all available information from just the audio and a focus on robustness.
I had a great time working on this with @deezer in Paris! Big thanks to my mentors @evpure, @Gabolsgabs, and @m_schedl!
💻 Code: github.com/deezer/robust-AI-…
📄 ISMIR Paper (foundation): arxiv.org/abs/2506.18488
📄 ACL Paper (Multi-View Double Entendre): aclanthology.org/2025.findin…
Wtpsplit, our text segmentation tool, just reached ⭐️1000 stars⭐️ on GitHub! Excited to see it is proving useful!
Check it out here: github.com/segment-any-text/… 🎉
