MADARA: New routing cuts multi-agent RAG costs with model-adaptive isolation

Multi-agent document assessment in RAG systems is computationally expensive, pushing practitioners toward smaller, deployable models. A new paper from Jungseob Lee, Chanjun Park, and Heuiseok Lim at Korea University investigates how these smaller models actually benefit from assessment, revealing a sharp dichotomy. On 7B-9B instruction-tuned models across diverse QA benchmarks, they found that for weaker baselines, the dominant mechanism is per-document isolation — not scoring quality. Astonishingly, assessment-free isolation alone matches full multi-agent assessment, showing that resolving multi-document context confusion drives outsized gains of up to 50 percentage points.

For stronger baselines where scoring quality does matter, the authors introduce Reasoning-Score Coupling, a label-free perturbation probe that classifies scoring behavior. Integrating these findings, they propose MADARA (Model-Adaptive Document Assessment Routing Architecture). Crucially, MADARA's diagnostic thresholds, derived from a single pilot model, generalize zero-shot to four unseen model families — providing a robust, lightweight pipeline to eliminate computational overhead. This work offers practical guidance for building cost-effective multi-agent RAG systems by adapting assessment strategies to model capability.