Researchers at the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) presented three papers at the International Conference on Learning Representations, indicating breakthroughs in Large Language Models’ (LLMs) abilities to form useful abstractions. The team used everyday words for context in code synthesis, AI planning, and robotic navigation and manipulation.
The three frameworks, LILO, Ada, and LGA, function as libraries of unique abstractions for particular tasks. LILO synthesizes, compresses, and documents code, Ada aids sequential decisions made by AI agents, and LGA enhances robots’ understanding of their environments.
LILO pairs a standard LLM with Stitch, the previously developed MIT refactoring tool, to write code and find abstractions. Refactoring trims down and combines code into libraries of reusable and readable programs. LILO’s use of natural language enables it to execute tasks requiring human-like knowledge, such as identifying and removing all vowels from a string of code and sketching a snowflake.
Ada, named after Ada Lovelace, is aimed at automating multistep tasks like virtual kitchen chores and gaming. It trains on possible tasks and their language descriptions, proposing action abstractions from this dataset. A human operator then scores and filters these into a library for application in hierarchical plans for different tasks.
LGA (language-guided abstraction) assists robots in interpreting their surroundings and understanding the essential elements needed to perform a given task. It can develop effective plans in unstructured environments, offering potential assistance for autonomous vehicles and factory and kitchen robots.
The researchers aspire to build on their findings to build more comprehensive understanding of AI models within these specific tasks, focusing on expanding LLMs’ abilities in executory tasks. The application of these study results could potentially see the introduction of software tools that can automate more complex manual tasks, in and out of the programmable domain. MIT CSAIL members are senior authors for each paper: Joshua Tenenbaum, Julie Shah, and Jacob Andreas.
