Researchers at Sakana AI have developed Text-to-LoRA (T2L), a hypernetwork designed to generate task-specific LoRA adapters for large language models (LLMs) using only natural language descriptions of tasks. Unlike traditional fine-tuning methods that require separate training for each downstream task, T2L can instantly create new LoRA adapters in a single forward pass by conditioning on a text prompt. The system also compresses many existing LoRAs into itself, enabling efficient adaptation of LLMs on the fly. This approach aims to reduce the complexity and computational expense associated with fine-tuning large models. The work was presented at ICML 2025 and authored by researchers including R. Charakorn, E. Cetin, Y. Tang, and R. T. Lange. Additional research highlights include advancements in LLM capabilities such as programming by backpropagation, where models trained solely on source code can execute programs without explicit input-output examples, and emerging strategies in in-context learning. These developments reflect ongoing efforts to improve the adaptability and efficiency of large-scale AI models.