MIT CSAIL’s Recursive Language Models (RLMs) address the limitations of large language models (LLMs) by enabling them to process extensive prompts without increased memory costs or unique training. RLMs treat long prompts as an external environment, allowing models to analyze and extract needed snippets efficiently. This method replicates computer data storage systems, where only relevant data is retrieved as needed. RLMs function via a Read-Eval-Print Loop (REPL), letting LLMs use coding commands to interact with prompts. They require advanced reasoning models like GPT-5 for optimal performance, while standard models may struggle. With a structure combining a root LLM and faster recursive LMs, RLMs can handle inputs far beyond conventional limits, effectively managing up to 10 million tokens. This innovation, aimed at enhancing memory management in LLMs, is set for integration into the DSPy framework and is available on GitHub, marking a significant evolution in AI language processing.
Source link
