When we put lots of text (eg a code repo) into LLM context, cost soars b/c of the KV cache’s size. What if we trained a smaller KV cache for our documents offline? Using a test-time training recipe we call self-study, we find that this can reduce cache memory on avg 39x

sleep-time compute* for kv cache *arxiv.org/abs/2504.13171

Even the smartest LLMs can fail at basic multiturn communication Ask for grocery help → without asking where you live 🤦‍♀️ Ask to write articles → assumes your preferences 🤷🏻‍♀️ ⭐️CollabLLM (top 1%; oral ICML Conference) transforms LLMs from passive responders into active collaborators.

Text diffusion models might be the most unintuitive architecture around Like: let's start randomly filling in words in a paragraph and iterate enough times to get something sensible But now that google's gemini diffusion is near sota, I think we need to take them seriously

How can we close the generation-verification gap when LLMs produce correct answers but fail to select them? 🧵 Introducing Weaver: a framework that combines multiple weak verifiers (reward models + LM judges) to achieve o3-mini-level accuracy with much cheaper non-reasoning

Let's go Minions! Avanika Narayan

One of today's key trends imo 👇 (By the way, the last point about delegation is why keeping an eye on projects like Minions is a good idea ...)

1/10 ML can solve PDEs – but precision🔬is still a challenge. Towards high-precision methods for scientific problems, we introduce BWLer 🎳, a new architecture for physics-informed learning achieving (near-)machine-precision (up to 10⁻¹² RMSE) on benchmark PDEs. 🧵How it works:

Was extremely fun to work on this paper with Jerry Liu and finally fulfilling our 7 year plan from year one of undergrad to write a paper together! One of many I hope!

This is awesome work from Jerry Liu and Yasa Baig "10 billion times better than prior work" to be exact

Agent swarm

Sure, local LLMs aren't great now, but you can use your imagination a bit and get creative with how you use them! - Cerebras and groq reduce latency which is great, but local compute has different advantages: privacy and utilization of compute that is otherwise sitting idle.

We're excited to announce a new research release from the Cartesia team, as part of a long-term collaboration to advance deep learning architectures. We've always believed that model architectures remain a fundamental bottleneck in building truly intelligent systems. H-Nets are

The concept of tokenization (or chunking more broadly) is fundamental to the way humans consume, process, and react to information. In language, vision, audio, haptics, and many more applications. Seeing that we can model such fundamental principles in an elegant,

At Cartesia, we've always believed that model architectures remain a fundamental bottleneck in building truly intelligent systems. Intelligence that can interact and reason over massive amounts of context over decade-long timescales. This research is an important step in our

Tokenizer free models! Deep learning has been a story of end-to-end learning replacing hand crafted features, so this next step feels fundamentally important (especially for modalities that are hard to tokenize like audio and DNA). Also lots of cool implications for multimodal

We’re thrilled to collaborate with the hazyresearch Stanford AI Lab, led by Chris Ré, to power Minions, their cutting-edge agentic framework tackling the cost-accuracy tradeoff in modern AI systems. This innovation is enabled on AMD Ryzen AI, thanks to seamless integration with

🎉 Excited to share that our paper "Pretrained Hybrids with MAD Skills" was accepted to Conference on Language Modeling 2025! We introduce Manticore - a framework for automatically creating hybrid LMs from pretrained models without training from scratch. 🧵[1/n]