Self-reflective program search improves recursive language model context use
July 8, 2026
Recursive Language Models (RLMs) use programmatic interaction to decompose long contexts, but suffer from inefficient trajectory selection. This work proposes self-reflective program search to improve the reliability of information extraction across extended context windows.
HOW THIS AFFECTS YOU
●
builderYou can use these agentic decomposition patterns to improve reasoning over very large documents.
●
researcherThis offers a way to improve RLM effectiveness through better trajectory selection.