Artificial intelligence, machine learning, computer science and mathematics.
An AI system designed for damaged South Indian palm-leaf manuscripts reported 97.4% word accuracy and 98.6% character accuracy on an archival corpus of 1,250 folios. It also reported 99.4% accuracy in identifying authentic palm-leaf manuscripts.
People can interpret the same general statement in different ways, yet generic language may still pass on broad knowledge. Across three experiments, a model in which such statements have an uncertain threshold best explained participants’ varied interpretations of new generic claims.
Researchers developed an attack that generates backdoor triggers tailored to a shared AI model in federated learning. In experiments across multiple datasets, it improved the attack success rate by 60.95–80.19% over average baseline performance without increasing the malicious updates sent to the system.
An analysis of GitHub Copilot use found that open-source projects produced more code and involved more developers. The gains came with longer coordination time, especially for developers on the edges of project communities.
A systematic literature review catalogs unresolved technical challenges in developing artificial general intelligence, including limited common sense, unreliable responses and difficulty handling unfamiliar situations. It points to hybrid systems, embodied approaches and stronger reasoning as possible research directions, while emphasizing the need for better safety and verification.
A new open-source pipeline uses an existing AI system to segment far more of very large Earth-observation images without training data or manual threshold tuning. Across seven scenes, it increased coverage to 91–98% and joined objects split across image tiles.
A new preprint reports a complete computer-assisted verification of the Lonely Runner Conjecture for 14 runners, with 13 of them moving. The archive includes certificates, source code and an audit script; because it is a preprint, peer review may change the findings.
A survey of human-AI collaboration argues that more capable AI systems do not automatically produce reliable partnerships. It identifies human-centered design, governance and attention to high-stakes uses as central to making these systems trustworthy and beneficial.
Deep-learning tools for predicting RNA splicing can rely on genomic clues unrelated to splicing and miss important effects of RNA structure, a study finds. These weaknesses were linked to systematic errors, especially when the tools evaluated sequences that differed from those used for training.
AI vulnerability detectors often perform less well on software projects outside their training data. A study finds that cleaner, more varied training data and an encoder-based design improved detection, including a reported 6.8% gain in recall on the BigVul benchmark.
Researchers developed a quantum method for the maximum cut problem and proved a stronger worst-case guarantee than a three-round version of a widely used approach. On IBM quantum devices with 72 and 148 qubits, the method exceeded a known classical hardness threshold, while the comparison method did not do so in the 148-qubit test.
A study finds that shared data from split-learning phishing detectors can reveal private webpage information. A new attack using large language models outperformed conventional reconstruction attacks in experiments on real-world phishing datasets.
A lightweight AI system detected marine organisms in degraded underwater images with lower computing demands than YOLOv8m. Tests on a dataset and an embedded computer suggest it could support faster, more energy-efficient monitoring on autonomous underwater vehicles.
A preprint study found that the same hidden instruction could produce very different results in AI agents depending on whether it arrived as tool information or tool output, and what surrounded it. The findings have not been peer reviewed and may change.
A marker-free augmented-reality system projected planned robotic-surgery port locations onto a human phantom. The results show consistent alignment between the digital plan and the physical body model, but the system has not yet been evaluated in patients.
