language detoxification

Recent studies have shown that large language models (LLMs), when augmented with additional weights through adapters—for instance, LoRA—can learn specialized tasks without updating model parameters—a paradigm known as either Instruction Tuning (IT) for individual instructions or Task Arithmetic (TA) for task-specific adaptation. In such scenarios, different instruction/task datasets induce distinct steering vectors within LLMs' embedding space, shaping behavior according to downstream objectives. Inspired by recent advances in model merging or parameter composition, we investigate whether multiple steering directions (each optimized toward generating responses on one dataset while discouraging outputs outside its distribution via unlearning-like objectives) can be combined into a single vector enabling zero-shot generalization across diverse tasks.