Beihang University researchers unveil a new federated learning poisoning-attack mitigation strategy to cap accuracy losses under 1% on MNIST and under 4% on CIFAR-10, bolstering AI security and privacy in healthcare, autonomous vehicles, and finance.

Researchers at the National University of Defense Technology and Shanghai Jiao Tong University introduce a symmetry-aware neural network architecture that slashes multi-agent reinforcement learning training time by up to 70%, doubles predator–prey task performance, and boosts sample-efficient coordination for autonomous traffic control, cooperative robotics, and power-grid management.

Researchers from Hefei University of Technology and Tsinghua University have developed a new cognitive diagnosis approach that overcomes student-concept sparsity to boost accuracy by up to 6%, delivering more accurate, real-time personalized learning analytics and equitable feedback.

Researchers at Nanjing University and China Mobile unveil a novel privacy-preserving traffic-engineering algorithm that cuts multi-domain routing time by 24.35% while maintaining 90% link utilization through differential privacy and graph neural network–driven bandwidth prediction.

A new multilingual AI survey by Beijing Foreign Studies University researchers uncovers uneven training data, cross-lingual alignment challenges, and embedded bias in 50+ large language models—offering a roadmap for balanced corpora, universal representation, and robust bias mitigation to boost low-resource language support.

A new semi-supervised audio analysis model leveraging bi-path feature extraction and attention mechanisms boosts sound event detection accuracy by 12.7% (PSDS1 0.489, PSDS2 0.771), enabling smarter smart-home, security, wildlife and industrial monitoring.