Tag
LatentSkill converts textual skills into LoRA adapters stored in weight space, reducing context overhead while maintaining modularity and composability for LLM agents, achieving significant improvements on ALFWorld and Search-QA benchmarks.
Proposes KOFF, a framework that decomposes pretrained LLMs into a sparse shared backbone and domain-specific external memories using structured pruning and LoRA adapters, achieving 12% sparsity without significant performance loss.
This paper characterizes backdoors in LoRA adapters that activate at the token feature level, and proposes behavioral and weight-level detection methods. The backdoor generalizes across related token patterns but not structurally identical ones, and detection methods show strong separation.
LoREnc is a training-free framework that secures foundation models and LoRA adapters via spectral truncation and compensation, preventing unauthorized model recovery while maintaining performance for authorized users. Accepted at ICIP 2026.