MIT researchers and Asari AI unveiled EnCompass, a framework that separates an AI agent’s workflow from its search strategy, enabling backtracking and parallel attempts when large language models err. In tests translating Java codebases to Python, EnCompass cut search-related coding by about 80% and, using a two-level beam search, improved accuracy 15% to 40% at a 16x LLM call budget. The system supports plug-and-play strategies such as beam search and Monte Carlo tree search, and targets agents with explicit programmatic workflows rather than fully free-form LLM control. The work, presented at NeurIPS and supported by Asari AI, aims to scale agentic tools to tasks like managing large code libraries and experiment design.
Related articles:
DSPy: Programming with Foundation Models
Tree of Thoughts: Deliberate Problem Solving with Large Language Models
Voyager: An Open-Ended Embodied Agent with Large Language Models
SWE-bench: Benchmarking LLMs on Real-World Software Issues
Self-Consistency Improves Chain-of-Thought Reasoning in Language Models
