To Isolate or to Score? Model-Adaptive Assessment for Cost-Efficient Multi-Agent RAG
06:00 · June 25, 2026 · arXiv cs.AI RSS
Multi-agent document assessment for retrieval-augmented generation is computationally expensive, driving practitioners toward smaller, deployable models whose assessment mechanisms remain poorly understood. We conduct a controlled study of training-free interventions on 7B-9B instruction-tuned models across diverse QA benchmarks, revealing a sharp dichotomy in how models benefit from assessment. For weaker baselines, the dominant mechanism is per-document isolation. Astoundingly, assessment-free isolation matches full multi-agent assessment, demonstrating that resolving multi-document context confusion, rather than scoring quality, drives outsized gains of up to 50 percentage points. Conversely, for strong baselines where scoring quality matters, we introduce Reasoning-Score Coupling, a label-free perturbation probe that classifies scoring behavior. Integrating these findings, we propose MADARA, a model-adaptive 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.
Summary
Multi-agent document assessment in retrieval-augmented generation multiplies inference cost by the product of rounds, agents, and documents, prompting practitioners to substitute 7B–9B instruction-tuned models for larger ones. A controlled ablation across QA benchmarks shows that these smaller models do not behave uniformly. Weaker baselines improve primarily through per-document isolation, which removes multi-document context confusion; even random, assessment-free isolation reproduces the full gains of multi-agent pipelines, reaching 25–36 points on adversarial tasks and up to 50 points on standard QA. Consequently, the expensive scoring and debate stages become redundant for these models.
Stronger baselines exhibit the opposite pattern: isolation yields no benefit, while the quality of the numerical scores themselves drives performance. To identify this regime without labels, the authors introduce Reasoning-Score Coupling, a perturbation probe that checks whether document scores degrade monotonically when reasoning quality is systematically impaired. The resulting diagnosis determines the appropriate lightweight intervention—per-document extraction for weak models or continued scoring with chain-of-thought adjustments for strong ones.
These observations are combined in MADARA, a routing architecture that first applies the probe and a no-filter baseline to classify each model–task pair, then selects the matching treatment. Thresholds calibrated on a single pilot model transfer zero-shot to four additional model families, confirming that the isolation–scoring asymmetry reflects an intrinsic capability divide rather than an artifact of any one system. The resulting pipeline eliminates unnecessary assessment calls while preserving peak accuracy across the examined range of 7B–9B models.
Why it matters
This research is highly relevant for Dutch AI researchers and practitioners developing cost-efficient RAG systems using smaller, deployable models (7B-9B). The proposed MADARA architecture offers actionable methods to reduce computational overhead, aligning well with the resource constraints and efficiency goals of Dutch SMEs.
