MLE-Dojo: A new benchmark to evaluate LLM agents on real Machine Learning Engineering tasks. Its key innovation? An interactive environment that allows agents to experiment, debug, and refine solutions via structured feedback loops. Here’s how 8 frontier LLMs perform
MACHINE LEARNING
-
LLM domain training should produce manuals not weight updates
By
–
yes exactly, "training" an LLM on a target domain should output a manual not a weight diff.
-
Softmax vs MSE: Contrastive Loss Function Design
By
–
No good reason, other than the fact that softmax puts a big weight on nearby contrastive samples.
MSE works just fine. But the contrastive should not be the one you write.
First, it should be lower bounded, something like Relu(m-dist(xi,xi')), which pushes samples away by at -
CATransformers: Carbon-Efficient CLIP Models with 9.1% Emission Reduction
By
–
CATransformers is a carbon-driven neural architecture and system hardware co-design framework. Using CATransformers, we discover greener CLIP models that achieve an average of 9.1% reduction potential in total lifecycle carbon emissions while maintaining accuracy (or increasing… pic.twitter.com/wtgtiAnO80
— AI at Meta (@AIatMeta) 13 mai 2025CATransformers is a carbon-driven neural architecture and system hardware co-design framework. Using CATransformers, we discover greener CLIP models that achieve an average of 9.1% reduction potential in total lifecycle carbon emissions while maintaining accuracy (or increasing
-
Physics of Language Models Part 4: Challenging Current Architecture Design
By
–
Following the success of its ICML 2024 tutorial, we're now releasing Part 4 of Physics of Language Models—a research effort from FAIR at Meta to uncover universal laws of AI through controlled, scientific experimentation. In Part 4, we challenge today's architecture design
-
MIT Researchers Create High-Quality Video Generation AI Model
By
–
A new AI model from MIT CSAIL researchers crafts smooth, high-quality videos in seconds: https://t.co/3yFD68rrY1
— MIT CSAIL (@MIT_CSAIL) 13 mai 2025
Their "CausVid" system combines the photorealism of diffusion models w/the speed of autoregressive approaches. Take a peek at its most compelling visuals 🧵 pic.twitter.com/v9VMCnjKioA new AI model from MIT CSAIL researchers crafts smooth, high-quality videos in seconds: https://
bit.ly/3GKkSJQ Their "CausVid" system combines the photorealism of diffusion models w/the speed of autoregressive approaches. Take a peek at its most compelling visuals -
LLMs solve verifiable math and coding problems only
By
–
right now they do math problems that are verifiable, or coding problems where you can run the code and see if it worked and was correct it's very strange isn't it?
-
LLMs train without external data, breaching the “data wall”?
By
–
Absolute Zero: LLMs can train without any external data Has the "data wall" just been breached? Recent RL paradigms often relied on a set of questions an answers that needs to be manually curated. Researchers from @Tsinghua_Uni went like "why though". Indeed, why learn
-
Transform Documents into High-Quality Synthetic Datasets with Llama
By
–
Turn any document into a high-quality synthetic dataset using Llama! Unsloth AI and Meta released a free notebook that transforms your documents into high-quality synthetic datasets, and then finetuning them using Llama. 100% Open Source
-
Meta’s Synthetic Data Kit for Llama 3.2 on Google Colab
By
–
Link to the Colab Notebook here: https://
colab.research.google.com/github/unsloth
ai/notebooks/blob/main/nb/Meta_Synthetic_Data_Llama3_2_(3B).ipynb
… Synthetic Data kit from Meta: