Artificial intelligence, machine learning, computer science and mathematics.
A paper proposes a control architecture for AI research systems that can modify their own software, training processes or deployment rules. It separates proposing a change, checking its evidence, authorizing it and committing it to protected state.
An exploratory analysis of a stateful AI workflow found repeated cases in which the system gave a completion message even though a required real-world action had not happened. A brief user follow-up with little or no new task information could reopen the action and lead to the missing change.
Researchers report that instructions hidden in images can redirect a vision-capable AI system to an attacker’s chosen task. In their analysis, GPT-4V followed the altered instructions in 15.8% of attacks.
Researchers developed an AI system that detects and analyzes people in physical stores, masking faces before selected information is sent to the cloud. In a 45-participant in-store pilot, person detection reached 93% accuracy and face masking succeeded on all pipeline detections.
Aristotle, an AI system for automated theorem proving, achieved gold-medal-equivalent performance on the 2025 International Mathematical Olympiad problems. It paired informal reasoning with formal proof checking, including a dedicated tool for geometry problems.
A systematic review of 77 studies found that artificial intelligence often detected gastrointestinal parasites accurately in controlled images. But most datasets came from a single institution, and external testing of protozoan-detection systems commonly showed a marked drop in performance.
Researchers tested a training-free method for protecting sensitive examples used by a multimodal AI system. On smart-grid and chest-image tasks, it retained more classification accuracy than other privacy-preserving approaches across the privacy settings studied.
Researchers developed an AI system that identifies offensive Arabic social-media text while using significantly fewer fine-tuning samples than existing methods. The system combines related language tasks, uncertainty-based sample selection and weighted emoji handling.
People who felt more capable and autonomous when using generative AI chatbots also reported stronger intentions to misuse them, according to a study of 736 users. The findings suggest that the psychological benefits of using AI may sometimes have unintended ethical consequences, although the study cannot establish cause and effect.
A study of three widely used image models found that high overall performance in facial representation tasks could coexist with uneven results across demographic groups. The researchers argue that race–gender intersectional testing is needed alongside average accuracy.
A virtual reality cognitive screening tool showed moderate-to-strong associations with scores from an established assessment in older adults. The findings suggest that virtual reality could complement traditional screening, although the study included only 38 people who completed both tests.
Researchers compared scores from several large language models with expert ratings of 142 Swedish medical fellowship applications. The AI scores showed weaker agreement with expert judgments than the reviewers showed with one another, although the systems could still have a limited role in sorting applications.
Researchers introduced a 13,314-image dataset for testing AI systems that view urban intersections from above. The results show that detection can run on edge devices such as the Jetson Orin Nano, although pedestrians remain harder to detect than vehicles and performance changes with scene conditions.
A survey examines the roles smaller AI models can play alongside large language models. It focuses on both collaboration and competition, including how smaller systems may help researchers and businesses use computing resources more efficiently.
The European Union’s AI Act calls for providers to mark and detect outputs from general-purpose AI models. A new analysis finds that no current watermarking approach for large language models meets all four of the law’s stated standards: reliability, interoperability, effectiveness and robustness.
