Presenting our WWW 2025 paper "Harmful Terms and Where to Find Them" at the NDSS poster session today!
Many T&Cs hide unfavorable financial terms—hidden fees, tricky refunds, forced arbitration. We use LLMs to analyze 1.8M T&C lines and build a tool that flags them in real-time.
🔍 Paper (arXiv): https://t.co/NPWrYlYd4d
💻 Code (GitHub): https://t.co/OZhZZKjylT
📊 Dataset (Hugging Face): https://t.co/ZMq8b864su
🔍 Read more in my blog post: https://t.co/yZ3flNWwR7
It feels like LLMs don’t just boost productivity; they augment my thinking. I firmly believe that LLMs are changing human cognitive processes: streamlining learning, accelerating problem-solving, and reshaping how we interact with information. Ah, what a time to be alive. (9/9)
When I encountered unfamiliar words (English isn’t my native language), I used an LLM-based etymology tool (that I also found on my X timeline) to analyze them (https://t.co/rkD0K0M6sf). (8/9)
Language Models Use Trigonometry to Do Addition
"We first discover that numbers are represented in these LLMs as a generalized helix, which is strongly causally implicated for the tasks of addition and subtraction, and is also causally relevant for integer division, multiplication, and modular arithmetic. We
then propose that LLMs compute addition by manipulating this generalized helix using the “Clock” algorithm: to solve a + b, the helices for a and b are manipulated to produce the a + b answer helix which is then read out to model logits. "
Excited to attend #FOCI2023 and #PETS2023 in person this week! @MustbeElisa will be presenting our paper on detecting DNS manipulation using TLS certificate information on Thursday, July 13! Come find us in Section 7B!
I came across Philipp Winter's Research Power Tools. I found it very insightful and applicable (also interesting and amusing to read). I put the digest of it on telegraph:
https://t.co/ixghwYQDc8