Systematic Ablation Reveals Hidden Failures in Multi-Agent AI for Science
Abstract
Abstract Retrieval-augmented generation (RAG) systems increasingly ground large language models (LLMs) in scientific literature, yet their design choices are typically made by convention rather than evidence. Here, we introduce a systematic ablation methodology for RAG pipeline design, validated through triple triangulation of deterministic ground-truth metrics, LLM judge scoring, and natural language inference. Across more than 36,000 evaluations spanning 200 scientific papers and 250 expert-curated questions, we find that self-correction without retrieval grounding degrades 79% of correct answers. Iterative retrieval saturates logistically with depth, capturing 80% of the improvement within seven turns. LLM judge faithfulness scoring also exhibits a construct validity failure, where answers rated perfectly faithful have only 21% of their claims actually grounded in sources. These findings, replicated across 3,300 code generation evaluations, demonstrate that reliable AI for science demands not just better models, but disciplined evaluation methodology that exposes failure modes invisible to any single metric.
Read the paper
Related studies
- Knowledge Management Perspective of Generative Artificial Intelligence
- Navigating Generative Artificial Intelligence Promises and Perils for Knowledge and Creative Work
- Characterizing the evolving research on enterprise content management
- Large language models and multimodal foundation models for precision oncology
- From Generative AI to Extended Reality: Multidisciplinary Perspectives on the Challenges, Opportunities, and Future of Educational Computing
- Generative Artificial Intelligence: A New Engine for Advancing Environmental Science and Engineering
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.