TRANSFORMER2 SELF ADAPTIVE LLMS

This research paper introduces Transformer2, a novel self-adaptive large language model (LLM) framework. Transformer2 uses Singular Value Fine-tuning (SVF), a parameter-efficient method, to train "expert" vectors for specific tasks using reinforcement learning. During inference, a two-pass mechanism dynamically combines these experts based on the input prompt, significantly improving efficiency and performance compared to methods like LoRA. Three adaptation strategies are presented, showing improved results with increased access to test-time information. The framework demonstrates versatility across different LLM architectures and modalities, including vision-language tasks.


paper - https://arxiv.org/pdf/2501.06252v1
subscribe - https://t.me/arxivdotorg

created with NotebookLM