No Test Cases, No Problem: Distillation-Driven Code Generation for Scientific Workflows

Researchers Siddeshwar Raghavan and Tanwi Mallick have introduced MOSAIC, a novel multi-agent framework that tackles a critical limitation in LLM-based code generation: the lack of I/O test cases in scientific workflows. Traditional multi-agent systems rely on execution feedback from test cases to iteratively improve code, but scientific problems often lack such test cases—generating them would require solving the problem itself. MOSAIC sidesteps this entirely by employing a student-teacher knowledge distillation framework that grounds code generation in domain-specific examples and structured problem decomposition, all without any training or fine-tuning.

To maintain coherence across the chained subproblems that scientific workflows demand, MOSAIC introduces a Consolidated Context Window (CCW) that ensures consistent reasoning across all agents. This mitigates the hallucinations that typically plague multi-agent systems when tackling complex, multi-step problems. On the SciCode benchmark, MOSAIC demonstrated significant improvements in accuracy, executability, and numerical precision over existing approaches, all while relying on lightweight, computationally efficient models. This makes it particularly valuable for resource-constrained research environments.