🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Excited to share our paper "Universal Sharpness Dynamics..." is accepted to #ICLR2025! Neural net training exhibits rich curvature (sharpness) dynamics (sharpness reduction, progressive sharpening, Edge of Stability)- but why?🤔 We show that a minimal model captures it all! 1/n

Drop by our ICLR workshop tomorrow on Building Trust in LLMs and LLM Applications! We’ll cover topics ranging from guardrails to explainability to regulation and more. We'll be in Hall 4 #6: building-trust-in-llms.github.io/iclr-workshop/…

Following release of our recent work, we have spent considerable time engaging with lmarena.ai over last week. The organizers had concerns about the correctness of our work on the reliability of chatbot arena rankings.

Despite the rise in combining LLM and DiT architectures for T2I synthesis, its design remains severely understudied. We explore several architectural choices that affect this design. We provide an open & reproducible training recipe that works at scale. This was done long ago

Tried steering with SAEs and found that not all features behave as expected? Check out our new preprint - "SAEs Are Good for Steering - If You Select the Right Features" 🧵

Zero-Shot Vision Encoder Grafting via LLM Surrogates "We construct small “surrogate models” that share the same embedding space and representation language as the large target LLM by directly inheriting its shallow layers. Vision encoders trained on the surrogate can then be

Confused about recent LLM RL results where models improve without any ground-truth signal? We were too. Until we looked at the reported numbers of the Pre-RL models and realized they were serverely underreported across papers. We compiled discrepancies in a blog below🧵👇

🤔 What if you gave an LLM thousands of random human-written paragraphs and told it to write something new -- while copying 90% of its output from those texts? 🧟 You get what we call a Frankentext! 💡 Frankentexts are surprisingly coherent and tough for AI detectors to flag.

Excited to share Video-Skill-CoT🎬🛠️– a new framework for domain-adaptive video reasoning with skill-aware Chain-of-Thought (CoT) supervision! ⚡️Key Highlights: ➡️ Automatically extracts domain-specific reasoning skills from questions and organizes them into a unified taxonomy,

Most papers discuss the hallucination problem in visual language models. In this paper, we present a framework to quantify both hallucination and omission problems in modern video LLMs. Both dataset and benchmarking code out!

Excited to share the results of my internship research with AI at Meta, as part of a larger world modeling release! What subtle shortcuts are VideoLLMs taking on spatio-temporal questions? And how can we instead curate shortcut-robust examples at a large-scale? Details 👇🔬