Think of a Large Language Model (LLM) like a brilliant scholar. To do their job well, they don’t just need their own brain; they need a good workspace—a desk with the right books, a filing cabinet that’s easy to navigate, and a clear set of instructions on how to process information. In the tech world, this “workspace” is called a harness.

Up until now, these harnesses have been built by human engineers through trial and error. While we have tools to automatically improve the AI’s “brain” (the model weights), the code that actually manages the AI’s information has remained stubbornly manual.

Meta-Harness automatically optimizes model harnesses — the code determining what to store, retrieve, and present to an LLM — surpassing hand-designed systems on text classification, math reasoning, and agentic coding.

To get the best out of AI, some users tell it to provide answers as if it were an expert. Others ask it to adopt a persona, such as a safety monitor, to guide its responses. However, this approach can sometimes hurt performance, according to a study available on the arXiv preprint server.

To see how well large language models (LLMs) behave when they are told to be someone else, researchers from the University of California ran a huge test using 12 different personas across six language models. These included experts in fields like math, coding and STEM (science, technology, engineering and mathematics) as well as general roles such as creative writer or safety monitor.

The team found that adopting a persona was something of a double-edged sword. While it makes AI sound more professional and keeps it safer (more likely to follow rules and less likely to generate harmful content), it sometimes performs worse at recalling facts.