Research
Papers, benchmarks and the techniques behind what eventually ships.
-
Gallup finds 59% of US teachers used AI last school year, and most say it freed up time
A Gallup survey of 1,952 US educators found 59% used AI in 2025-26, and about two-thirds of them said it gave them…
-
Robot failure detectors have less room for false alarms than their scores suggest
A new framework re-scores robot failure monitors by what it costs to act on them, and finds that even an ideal detector…
-
A humanoid policy trained almost entirely on human video
A European team reports a humanoid control policy trained on 93 per cent human motion data that follows walking instructions on a…
-
IIIT Delhi finds maternal-health models match the average mother, not the individual
An IIIT Delhi study of ten models on maternal-health advice finds they track the average north Indian mother closely but barely shift…
-
FAR.AI reports internal probes catch model sabotage that text monitoring misses
FAR.AI reports that probes reading a model’s internal activations beat text-based monitoring at spotting sabotage, on the authors’ own benchmark numbers.
-
LimX Dynamics trains a humanoid model to predict what the robot actually did
A Shenzhen humanoid maker reports better manipulation by training its model to predict the robot’s own body state after a move, not…
-
Robot reinforcement learning at a million environments stalls on reset sampling, not compute
A CoRL 2026 paper argues that very large-scale robot reinforcement learning is held back by which starting positions the simulator serves up,…
-
Ai2 rebuilt GPU scheduling around time budgets, and the gains are in queue waits
Ai2 has published how it replaced priority queues on its research GPU clusters with allocated time budgets, and the measurable gains are…
-
Anthropic puts $150m behind US federal science projects
Anthropic has committed $150 million over three years to the US Genesis Mission, without saying how much of it is cash rather…
-
A gene-finding model and a database of 566 million predicted loci
Hugging Face’s biology group has released a model that predicts protein-coding regions straight from DNA, plus annotations across more than 22,000 species.