🔥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

Parameterized Transforms 🚀 Here is a new tool that provides a modular and extendable implementation of torchvision-based image augmentations that provides access to their parameterization. [1/5]

"A calculator app? Anyone could make that." Not true. A calculator should show you the result of the mathematical expression you entered. That's much, much harder than it sounds. What I'm about to tell you is the greatest calculator app development story ever told.

Can machine learning models predict their own errors 🤯 ? In a new preprint w/ Apple collaborators Aravind Gollakota, Parikshit Gopalan, Charlotte Peale, and Udi Wieder, we present a theory of loss prediction and show an equivalence with algorithmic fairness! A thread (1/n):

I’m excited to announce that 💫StarVector has been accepted at CVPR 2025! Over a year in the making, StarVector opens a new paradigm for Scalable Vector Graphics (SVG) generation by harnessing multimodal LLMs to generate SVG code that aesthetically mirrors input images and text.

StarVector is out on Hugging Face StarVector is a foundation model for generating Scalable Vector Graphics (SVG) code from images and text. It utilizes a Vision-Language Modeling architecture to understand both visual and textual inputs, enabling high-quality vectorization

You can try the official StarVector demo in Hugging Face Spaces: 1B🔗huggingface.co/spaces/starvec… 8B🔗huggingface.co/spaces/starvec…

Following the excitement around our recent 💫 StarVector work, we're thrilled to introduce our latest research on benchmarking UI Vision Agents. We present UI-Vision, a benchmark designed to evaluate multimodal LLMs in their UI perception, spatial reasoning, and interaction

rizal˙ Jeremy Blaze StarVector does a very decent job on this example! Try here: huggingface.co/spaces/starvec…

Scaling Laws for Native Multimodal Models - Early-fusion exhibits stronger perf at lower param counts, is more efficient to train, and is easier to deploy, compared w/ late fusion. - Incorporating MoEs allows for models that learn modality-specific weights, significantly

Our work on fine-grained control of LLMs and diffusion models via Activation Transport will be presented ICLR 2025 as spotlight✨Check out our new blog post machinelearning.apple.com/research/trans…

Apparently you can trick an LLM to believe it did its thinking, and results improve by a lot (up to 40%)! In what authors call NoThinking they just force the string "Thinking: Okay, I think I have finished thinking" From paper: Reasoning Models Can Be Effective Without Thinking

Many design choices in (self-)supervised classification lead to identifiability. If you are curious how and when optimizing cross entropy leads to theoretical guarantees, check out our oral on Saturday in Session 5C or the poster from 3pm. Details: iclr.cc/virtual/2025/o…

Stop by poster #596 at 10A-1230P tomorrow (Fri 25 April) at #ICLR2025 to hear more about Sigmoid Attention! We just pushed 8 trajectory checkpoints each for two 7B LLMs for Sigmoid Attention and a 1:1 Softmax Attention (trained with a deterministic dataloader for 1T tokens): -

If you are interested in training multimodal models, this paper from my colleagues Apple is a gem!

This is awesome Juan A. Rodríguez 💫!

Is input selectivity the secret sauce of Mamba or there is more🤔? In our new ICML2025 paper, we show input selectivity provides an edge but conv and gating are key for associative recall! Check out Teresa Huang 's thread for more insights! Paper: arxiv.org/abs/2506.11891

Come check out our poster today at the East Exhibition Hall (E-2312) from 11am-130pm PDT!

Apple Intelligence Foundation Language Models: Tech Report 2025 "We introduce two multilingual, multimodal foundation language models that power Apple Intelligence features across Apple devices and services: i a 3B-parameter on-device model optimized for Apple silicon through

Uncertainty methods and correctness metrics often share "mutual bias" (systematic errors from a common confounder like response length), skewing LLM evaluations. New paper from my colleagues shows that "LM-as-a-judge" evaluation is more robust and human-aligned. Important work -

When you go back home and find your parents raise their monitor with your PhD thesis 😅