This article introduces Transformer², a new adaptive framework proposed by Sakana AI, which solves the computationally intensive and static shortcomings of traditional large language model (LLM) fine-tuning methods. Transformer² uses a two-stage mechanism to adjust LLM weights in real time during the reasoning process, allowing it to flexibly adapt to various unknown tasks and adapt to the environment like an octopus. Its core lies in singular value fine-tuning (SVF) and adaptive strategies, which train "expert" vectors through reinforcement learning and dynamically combine these vectors to achieve accurate responses to different tasks. This framework has many advantages such as parameter efficiency, modularity, and cross-model compatibility, and has demonstrated better performance than traditional fine-tuning methods in experiments.
The core of Transformer² is its unique two-stage mechanism and singular value fine-tuning (SVF) technology, as well as the clever combination of multiple adaptive strategies. The "expert" vectors trained through reinforcement learning give the model strong adaptability, allowing it to perform well in a variety of unknown tasks. Although there is still room for improvement, Transformer² has undoubtedly taken an important step towards building a truly dynamic, self-organizing AI system. Future research directions include model merging and the expansion of CEM methods. The address of the paper is attached at the end of the article, and we look forward to more researchers exploring this in depth.