The COSTA system pioneers cascade spatio-temporal attention plus memory-distilled iterative cross-modal alignment, cutting redundant video clutter and lifting dialogue response quality to new SOTA on long, dynamic videos.

SGR turns adversarial shortcuts into teachers, isolating invariant subgraph rationales and delivering the first XGL framework provably robust against spurious correlations on both synthetic and real-world graphs.

Research team debuts the first deterministic streaming algorithms for non-monotone submodular maximization, delivering superior approximation ratios with minimal memory and real-time throughput on massive datasets.

 Research team releases the first panoramic LoRA survey, mapping five research fronts—task adaptation, plug-in generalization, speed-ups, federated privacy and real-world deployment—and charting GaaS, continued pre-training and agent roles as next horizons.

Researchers deliver the first comprehensive survey and open NNLCB benchmark of neural universal lossless compressors, taxonomizing 17 state-of-the-art static/adaptive/semi-adaptive algorithms tested across 8.5 GB of multi-source data on CPUs/GPUs.