Reproduce “Sharing State Between Prompts and Programs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ellie Y. Cheng, Logan Weber, Tian Jin, Michael Carbin
Researchers at MIT CSAIL introduced the "shared program state" abstraction, allowing natural language code to directly interact with a formal host program's variables, objects, and control flow. This approach, implemented in the NIGHTJAR system, reduces boilerplate code by 39.6% and achieves comparable or higher task accuracy for hybrid natural-formal programs, albeit with increased runtime overhead.
Key insights
- Direct manipulation of program state via effects and handlers offers a more integrated programming paradigm, going beyond traditional function calling or tool use for LLMs.
- Host-language specialization and system engineering, such as eager variable loading and efficient reference handling for mutable objects, are crucial for practical performance and feature support.
- The shared program state significantly improves program conciseness and enables handling of large data structures by selectively querying necessary parts, overcoming context window limitations of pass-by-copy methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.