Ayal Klein

- Beyond Surface Level: The model isn't just matching patterns; it has internalized informal register as a portable, language-agnostic pragmatic abstraction. Basically, we found the "slang" dial in the LLM's brain

What we found: - The "Informal Register Subspace": We identified a geometrically coherent "island" of features shared across English, Hebrew, and Russian. - Zero-Shot Steering: Activating these features causally shift the model's output formality zero-shot across unseen languages

Does an LLM see "slang" as just a bunch of memorized words, or does it actually understand the abstract concept of being informal? In our latest work, we dive into the internal "mind" of Gemma-2-9B using Sparse Autoencoders (SAEs) to find out. 🧠✨

CLPsych 2025 shared task registration is now open -- deadline is February 9th. We hope you can participate! clpsych.org/shared-task/

Equal contribution @meetdavidwan Collaboration @ArieCattan, @kleinay2, Ido Dagan, and @mohitban47 This work was a great collaboration between @uncnlp and @biunlp! 💻 Check out our code here: github.com/ZhangShiyue/QA…

🚀 New paper out! 📜 Tackling a long-standing challenge in text summarization: How do we properly conduct human evaluations to improve its systematicity, granularity, and scalability? 🤔 Let’s dive in. 🧵👇 📖 Read the full paper here: arxiv.org/abs/2412.07096

check out the live demo for our joint QA-SRL - QANom model at huggingface: huggingface.co/spaces/kleinay…

We release the first unified QASem parsing tool, providing a QA-based systematic representation of sentence meaning. QASem covers pred-arg relations for verbs and nominalizations + discourse relations. Practical as an explicit account of the information units in a text.

Excited to share our work on QASem Parsing, a text-to-text modeling of QA-based semantics! 🤩 Preprint: arxiv.org/abs/2205.11413 Code: github.com/kleinay/QASem @hirscheran @EliavRon @valentina__py @clu_avi IdoDagan

Written texts consist of multitudes of nominalizations. In our new paper at @coling2020, we propose to capture the semantic roles of their arguments using NL questions (QANom), extending QASRL beyond verbs. Check it out- aclweb.org/anthology/2020…