TL;DR: We study the privacy risks of partial model merging and show that private model components can still be substantially reconstructed under realistic clone-attack settings. 

TL;DR: We expose a new poisoning vulnerability in multimodal LLM recommender systems, showing that coordinated text–image perturbations can reliably promote target items while preserving normal recommendation utility. 

TL;DR: We propose an inference-time pruning framework that maps domain-aligned representation subspaces to sparse model pathways, enabling scenario-specific LLM subnetworks with improved efficiency. 

TL;DR: We introduce XDomainBench, a benchmark for interactive interdisciplinary scientific reasoning, and show that LLMs exhibit systematic reasoning collapse as cross-domain composition becomes more complex. 

TL;DR: We introduce FedAdamom, which adapts server momentum instead of the learning rate, to keep fast saddle-point escape while better selecting flat minima, improving both convergence and test accuracy.

TL;DR: A lightweight framework that integrates federated learning and targeted unlearning for LLMs, formulating them as a joint multi-objective optimization task to enable privacy-preserving training and compliance with GDPR's right to be forgotten. 

TL;DR: A metric using limited unlabeled data to measure how well models can be merged by comparing merged weights vs ensemble outputs at layer/node levels. It guides merging and pruning and improves alignment with ensembling.