Taxonomy of Weight, Prompt, and Embedding-Based LLM Task Adaptation
October 2, 2026
This survey provides a unified framework for task adaptation, categorizing methods by where task information is encoded: model weights, input prompts, or injected embeddings. It analyzes the trade-offs between parameter-efficient fine-tuning, in-context learning, and emerging embedding-based approaches.
HOW THIS AFFECTS YOU
●
builderUse this to select the most efficient adaptation method based on your encoding requirements.
●
researcherYou can use this taxonomy to navigate the cross-paradigm relationships between PEFT and prompting.