Apple researchers compare two ways for LLM agents to find the right tools in large registries: retrieving tools similar to the user query, or first having the model sketch the function signatures it needs and searching with those. The question matters because context windows cannot hold every tool definition.
Apple's machine learning team describes how to run frontier-scale language models such as Muse Spark Open locally and across several machines. The setup relies on its MLX framework and RDMA over Thunderbolt 5 on Mac Studio computers with the M5 Ultra chip.
Apple researchers propose Selection-based Structured Reasoning (SSR), in which a multimodal agent picks from a set of predefined reasoning steps instead of writing free-form reasoning before each action. The aim is to cut the inference cost of long, unhelpful reasoning in small models.
Datalab has released OmniParseBench, an open benchmark of 16,288 yes-or-no tests spanning 2,937 pages in 92 languages. Each score is a pass rate that can be broken down by tag, script or language, making it easier to see where OCR parsers fail.
The founders of Ricursive Intelligence describe AI that designs chips, learns from each iteration and improves the next design. Their goal is to automate and speed up chip design to ease the shortage of computing capacity.
In a Latent Space interview, Leash Bio's Ian Quigley explains how the company mastered DNA-encoded library assays, a notoriously difficult technique, and used the resulting data to detect models that cheat. He also argues the same data is well suited to training AI agents to reason like chemists.