This thesis introduces two novel approaches for Temporal Action Localization (TAL) in video understanding. The Long-Short-range Adapter (LoSA) is a memory-efficient backbone adapter for untrimmed videos, modifying intermediate layers across various temporal ranges to enhance video features. It enables end-to-end adaptation of billion-parameter models like VideoMAEv2. The OVFormer framework addresses Open-Vocabulary TAL by generating rich class descriptions using a language model, aligning these with video features through cross-attention, and employing a two-stage training strategy for novel category generalization. LoSA enables efficient use of state-of-the-art video models, while OVFormer expands recognizable actions beyond predefined categories. These contributions significantly advance TAL, enhancing both capability and flexibility in action recognition and paving the way for more versatile video understanding systems.