DAST: Context-Aware Compression in LLMs via Dynamic Allocation of Soft Tokens

Open in new window