We're posting this preprint alongside another Hasson Lab manuscript led by Zaid Zada and Sam Nastase where we use large language models to capture brain-to-brain linguistic coupling during dyadic conversations—check it out: doi.org/10.1101/2023.0… x.com/samnastase/sta…

זה הסלפי של אחרי העברת חוק הסבירות. באותה שעה בכנסת עמדו אלופים ביחד עם הרמטכ״ל בכנסת לאחר שהתחננו לדבר איתם, להסביר להם שמדובר בנזק בטחוני עצום. הם לא הסכימו אפילו לשמוע אותם. תראו את החיוכים של האנשים האלה שהביא למחדל של למעלה מ1600 נרצחים וחטופים. ארורים אתם, ארורים לעולם

פנייה ממני לכל אזרחי ישראל.

In a new version of our (Mariano Schain, Ariel Goldstein, Hasson Lab) paper (arxiv.org/abs/2310.07106) we show evidence that the layered hierarchy of large language models (LLMs) may be used to model the temporal dynamics of language comprehension in the brain...

1/9: Our recent study, Decoding Stumpers: Large Language Models vs. Human Problem-Solvers, was published in #EMNLP2023. Alon Goldstein, Miriam Havin , Roi Reichart & Ariel Goldstein

How do the inherent biases in Large Language Models impact their capacity for accurately simulating human behavior? We (yanivdover, Roi Reichart, Ariel Goldstein) try to answer this question in our new paper arxiv.org/abs/2402.04049 1/9

🧠 Can LLMs truly understand text, or are they mere 🦜? We argue that this debate implicitly revolves around the function of consciousness, an age-old question which almost by definition cannot be answered via objective measurement. arxiv.org/pdf/2403.00499… w/Ariel Goldstein

To demonstrate this, imagine a “zombie” chatbot Z. It is completely open-source (à la OLMo), and it excels in all current and future “understanding” benchmarks, devoid of any subjective experience. Would you consider it intelligent or capable of understanding?

We show that seminal works in AI would likely consider Z incapable of understanding because, regardless of its external success, it still lacks internal experience. Instead, we propose to disentangle “understanding”, “intelligence”, “empathy”, and other human traits that...

supposedly require consciousness, from functional understanding, which measures objective success on given instances. This distinction yields vastly different research agendas: working on functional understanding implies hill-climbing on increasingly challenging...

datasets and tasks, while research in machine consciousness may aim to imbue machines with internal experience or to measure whether it somehow emerged in them.

🧠🤖 Does learning in the brain inherently require plasticity? In our latest paper we question this assumption, by leveraging insights into how LLMs "learn". Check out this thread for more details! biorxiv.org/content/10.110… w/ Yuval Shalev Gabriel Stanovsky Ariel Goldstein

Chreck out our new LLM-based suicide research paper at Frontiers - Psychiatry !

🧠🤖 How do LLMs think? What kind of thought processes can emerge from artificial intelligence? Our latest paper about multi-hop reasoning tasks reveals some new interesting insights. Check out this thread for more details! arxiv.org/abs/2406.13858 Ariel Goldstein Amir Feder

📢Paper release📢 What computation is the Transformer performing in the layers after the top-1 becomes fixed (a so called "saturation event")? We show that the next highest-ranked tokens also undergo saturation *in order* of their ranking. Preprint: arxiv.org/abs/2410.20210 1/4

Our findings reveal that this is true across language, vision, and speech models and across architecture variants: decoder-only, encoder-only, and full-Transformer. It even occurs in untrained Transformer models! 🤯 2/4

We propose an underlying task transition mechanism where each task corresponds to determining the k-th ranking token. By understanding these transitions, we can predict the current task from hidden layers representations and cause the model to switch between tasks! 3/4

🎯 Finally, we leverage these insights to introduce a new token-level early-exit strategy that beats existing methods in balancing performance and efficiency. More accurate predictions and faster models—win-win! Joint work with Ariel Goldstein Tomer Gabriel Stanovsky 4/4

Very excited to share our new paper published in Nature Communications Nature Communications (link below). This work is part of my PhD research under the supervision of Roi Reichart (Technion), @HassonUri (Hasson Lab), and Ariel Goldstein, in collaboration with Yoav Meiri🇮🇱.

New Preprint 🎉 LLM self-assessment unlocks efficient decoding ✅ Our Confidence-Informed Self-Consistency (CISC) method cuts compute without losing accuracy. We also rethink confidence evaluation & contribute to the debate on self-verification. arxiv.org/abs/2502.06233 1/8👇