MOSAIC framework enables multi-agent LLM code generation for scientific workflows without test cases

Researchers introduced MOSAIC, a training-free multi-agent framework for scientific code generation that does not require Input/Output test cases or execution feedback.

Instead of test-case-driven iteration, MOSAIC uses student-teacher knowledge distillation (a technique where a smaller model learns from a larger one) grounded in domain-specific examples and structured problem decomposition, and introduces a Consolidated Context Window to maintain consistent reasoning across agents.

On the SciCode benchmark, MOSAIC improved accuracy, executability, and numerical precision over existing approaches while relying on lightweight models.