To Isolate or to Score? Model-Adaptive Assessment for Cost-Efficient Multi-Agent RAG

Short summary

Research reveals weaker models benefit primarily from per-document isolation rather than scoring quality in multi-agent RAG (up to 50-point gains), while stronger models depend on assessment accuracy. MADARA, a new model-adaptive routing architecture, selects the appropriate strategy per model and generalizes zero-shot across different model families. This provides a lightweight pipeline to eliminate RAG computational overhead.

• •Document isolation drives gains for weaker models, not scoring quality
• •Stronger models rely on assessment accuracy to improve performance
• •MADARA routing architecture adapts per model with zero-shot generalization

Generated with AI, which can make mistakes.