Research to watch
Papers and lab posts worth knowing about, from research labs and the most-upvoted papers of the week.
Updated 24 Oct, 7 PM ET
Google shares how its AI helps screen for tuberculosis, from clinics to mines (opens the original source in a new tab)
Google says the chest X-ray screening tool that grew from its research is used in about 40 health facilities in six countries, and that miners are checked on site four kilometres underground.
Google Research · Original source (opens in a new tab)
Microsoft's Flint-interactive lets AI agents make charts you can click and explore (opens the original source in a new tab)
Microsoft Research says the language lets people and agents describe what a click or drag should do, such as highlighting linked points, while its compiler takes care of the low-level chart code.
Microsoft Research · Original source (opens in a new tab)
Microsoft AI publishes results for MAI-DxO, its medical diagnosis system, in Science (opens the original source in a new tab)Microsoft says MAI-DxO was 1.7 times more accurate than doctors on common emergency cases and more than twice as accurate on very hard journal cases, while using fewer tests.
Microsoft AI · Original source (opens in a new tab)
Google DeepMind and partners release a world map of forest types (opens the original source in a new tab)
ForTy 2020 sorts every 10-by-10-metre patch of land into six forest and land-use types, so you can tell natural forest from plantations, with an early record of yearly changes from 2017 to 2025.
Google DeepMind · Original source (opens in a new tab)
A Microsoft researcher and a collaborator use AI agents to help solve a 31-year-old maths problem (opens the original source in a new tab)
The KLS conjecture, a central question in high-dimensional geometry, is now resolved. AI systems explored ideas while the researchers judged and steered them, and other experts and a machine-checked proof confirmed it.
Microsoft Research · Original source (opens in a new tab)
Google's AI reads simple ultrasound sweeps to tell how far along a pregnancy is (opens the original source in a new tab)
In a study across several clinics, health workers without scan training moved low-cost probes over the belly in a set pattern, and the AI judged the pregnancy's age as well as experts, Google says.
Google Research · Original source (opens in a new tab)
Anthropic uses Claude to lock away code that attackers don't need to reach (opens the original source in a new tab)
Instead of hunting single bugs, Claude finds code that can be reached but isn't needed and moves it behind a gate, such as a login. Run over 25 open-source projects, it found 493 issues, Anthropic says.
Anthropic · Original source (opens in a new tab)
Claude designs proteins that latch onto scorpion toxins (opens the original source in a new tab)
Anthropic says an internal Claude model worked mostly on its own for 64 hours; of 211 designs tested in labs, 17 attached to dangerous scorpion toxins. Anthropic does not claim it has made an antivenom.
Anthropic · Original source (opens in a new tab)
Popular papers this week
The most-upvoted papers on Hugging Face Daily Papers, most upvotes first.
StreamEval tests video generators that must change a video while it keeps playing (opens the original source in a new tab)
The benchmark checks whether models can follow requests that arrive bit by bit during a running video stream, make the asked-for changes and keep the rest of the scene the same.
430 upvotes
FocusGUI helps app-using AI agents look only at what matters for each step (opens the original source in a new tab)
Agents that work apps take in screenshots, audio and video at once and get distracted. A small add-on picks the useful inputs for each action and, on average, tripled exact-match accuracy over the unchanged model.
222 upvotes