Very happy to share that this work got accepted to #NeurIPS2023 as a spotlight 🥳 It's my personal first ever acceptance at NeurIPS - and got an additional poster as cherry on top!

Due to their great success, LLMs have been increasingly used for scientific prediction and for uncovering the mechanisms behind scientific phenomena. This is true particularly when language is part of the mechanism or when it provides important signals, e.g. in fields like

1/15 📣preprint📣 TL;DR We (Yair Gat Amir Feder Alex Chapanin Amit Sharma Roi Reichart) show (theoretically and empirically) that #LLM-generated counterfactuals produce faithful SOTA explanations of how high-level concepts impact #NLP model predictions! arxiv.org/abs/2310.00603

1/🧵 Excited to share #Decipher 🔍, a game-changing method for integrating #singlecell RNA-seq data 🧬 from multiple conditions and revealing cell-state transitions in diseases like #AML. Dive into our thread for more! Check our preprint for full details. biorxiv.org/content/10.110…

Attending NeurIPS Conference #NeurIPS2023 next week? Join us for an enthralling discussion with Max Katz (from Martin Heinrich’s office), Zachary Lipton, Hoda Heidari, Katherine Lee & the incredible Louise Matsakis on how researchers can better help policymakers when it comes to

Super excited to present this work as spotlight tomorrow (Wed) at #NeurIPS23 alongside Claudia Shi & Amir Feder 🗓️10:45 - 12:45, Poster: #1523

Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations? New preprint!📣 - LLMs struggle to integrate new factual knowledge through fine-tuning - As the model eventually learns new knowledge, it becomes more prone to hallucinations😵‍💫 📜arxiv.org/pdf/2405.05904 🧵1/12👇

🧠🤖 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 out in final form: exciting results from our semi-supervised pose estimation package, Lightning Pose, which is now adopted by a number of great neuroscience labs. Please give it a whirl: github.com/danbider/light…

At #EMNLP2024? Join me in the Language Modeling 1 session tomorrow, 11:00-11:15, for a talk on how fine-tuning with new knowledge impacts hallucinations.

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👇

Excited for this work to be out 😀 Self consistency is great but v expensive (especially when you care about those last few acc points). We show: switching to a *weighted* majority vote (weights = confidence scores derived by the model itself) is way more sample efficient! 1/n

How can we use small LLMs to shift more AI workloads onto our laptops and phones? In our paper and open-source code, we pair on-device LLMs (ollama) with frontier LLMs in the cloud (@openai, @together), to solve token-intensive workloads on your 💻 at 17.5% of the cloud cost

📢 Excited to announce the #ICML2025 workshop on *Scaling Up Intervention Models (SIM)*! Let’s bring together state-of-the-art ideas on modeling novel interventions and distribution shifts. :) 🙌🏻 Submissions are welcome! Link: sites.google.com/view/sim-icml2…

This is going to be great! Looking forward to seeing you all at ICML.

new version of paper on worker rights in union contracts using NLP, with empirical work showing that these rights are valued like non-wage amenities. thread here: x.com/BenjaminArold/…

🚨 It's often claimed that LLMs know more facts than they show in their outputs, but what does this actually mean, and how can we measure this “hidden knowledge”? In our new paper, we clearly define this concept and design controlled experiments to test it. 1/🧵

⌨️ 😇 Drafting for NeurIPS? Submit to #ICML2025 workshop on Scaling Up Intervention Models (SIM) too! Let’s enjoy some fun science in Vancouver this July. 🌞🌳 🗓️Workshop submission due on May 20 AOE 🔗More info: sites.google.com/view/sim-icml2…

We secure all communications with a cloud-hosted LLM, running on an H100 in confidential mode. Latency overhead goes away once you cross the 10B model size. This is our first foray into applied cryptography -- help us refine our ideas.

Semantics in language is naturally hierarchical, but attempts to interpret LLMs often ignore this. Turns out: baking semantic hierarchy into sparse autoencoders can give big jumps in interpretability and efficiency. Thread + bonus musings on the value of SAEs: