צולף בכסף שהוא קל
Israel
Joined May 2016
- Tweets399
- Following1.9K
- Followers139
- Likes55.3K
צלף כסף קל retweeted
New paper:
We find reasoning models' CoT controllability to be much higher than current evals suggest.
Models rarely follow user instructions about their CoT, but follow them far more often when written as self-instructions inside the CoT, which can be used to evade monitors 🧵
צלף כסף קל retweeted
when you in a conversation w someone and they wanna connect deeply (u gotta bring it back to surface level shi)
32 מנדטים ב-2022 = 1.16 מיליון מצביעים.
26 מנדטים עכשיו = 1.17 מיליון מצביעים.
גם אחרי הארבע שנים הכי נוראיות בתולדות המדינה ביבי לא איבד אף מצביע. לא אכפת להם כמה דיממנו כל הקדנציה הזו.
החרא הזה באמת מדכא אותי ברמות
צלף כסף קל retweeted
מטרה אחת: להעיף את האיש שניסה לחסל את המדינה.
הכל כשר, כל חבירה שהגוש יחשוב עליה מערבים עד חרדים כשרה למען המטרה הזו.
אף אחד פה לא רוצה את ברדוגו שר משפטים!
תנו גב מלא לגוש להקים קואליציה עם מי שבא להם זולת הנאשם. נקים ונשבור את הראש לאחר מכן.
תנו להם גב, שחררו מהצדקנות לרגע קצר.
המסקנה היחידה מתוצאות האמת (הלוואי עוד ישתנו) היא שחייבים ממשלה גוש השינוי יחד עם כל המפלגות הערביות. יש 61 לזה.
צלף כסף קל retweeted
LLMs assist writers, yet they often sideline the human author as a passive participant.
Our new paper asks: what if the agent led the writing process the way expert human co-authors do, while keeping the human author actually engaged all along?
Meet CoWriter 🧵 (1/7)
ALT CoWriter agentic scheme overview: shared document, chat and sources on the left; a Leader agent guiding seven writing activities, a Monitor and a Reviewer agent on the right.
צלף כסף קל retweeted
Interesting finding about reasoning models and CoT.
They find that reasoning models can control their chain of thought much better than earlier tests suggested.
Earlier evaluations put the constraint in the user prompt and found very low controllability. Researchers put the same constraint inside the chain of thought as a self-instruction, and models followed it far more often.
A generic reminder to follow the user's constraint also helped a lot. The self-instruction can be triggered from the user prompt alone, so no access to the reasoning channel is needed.
Paper: arxiv.org/abs/2610.23257
Chat with Paper: academy.dair.ai/papers/reaso…
צלף כסף קל retweeted
“A language is a dialect with an army...,” goes the saying about Yiddish.
For LMs, training data may be that army.
When native texts are overshadowed by noisy translations, multilingual LMs can look fluent while missing what makes a language unique.
MameLoshnLM. #COLM2026 🧵
צלף כסף קל retweeted
In deep research reports, many sentences aren't supported by their citations: recall is 58.7% in NVIDIA AI-Q and 7.1% in TrajectoryKit.
Our #EMNLP2026 paper (main conf) proposes an algorithm that traces each error to the agent that introduced it, and identifies the error type.
צלף כסף קל retweeted
🚨new ACL paper🚨
🗄️dataset: LocQA🗄️
As we make models multilingual, do we inadvertently make them conflate knowledge across different languages? If we ask a model “who was the 1st president?” will it answer “George Washington” regardless of the language or locale? A 🧵
צלף כסף קל retweeted
What's in a neuron? 💫 (an atypically long, almost personal post)
Neurons in LMs have always been a fascinating object to study. I've been studying them since 2020, viewing them as key-value memory cells, analyzing what they capture in vocabulary space, and how they compose together to form features.
arxiv.org/abs/2012.14913
arxiv.org/abs/2203.14680
arxiv.org/abs/2501.08319
arxiv.org/abs/2506.10920
But many neurons still remain opaque! They do many things.
Our recent work led by @AsafAvrahamy tackles this challenge by decomposing neuron weights in vocabulary space. We do this by taking the neuron weight vector and learning different ways to rotate it (just a bit) to reveal monosemantic vocabulary channels that it captures. The nice thing about our method ROTATE is that it's data-free and super efficient, relying only on vocabulary kurtosis as a search signal.
I've been thinking about this idea since 2024, proposed it to multiple students, but only Asaf was brave enough to take this ;)
Very happy with the final outcome. Check out the paper! 👇
arxiv.org/abs/2604.06005
