A team of six researchers published a Systematization of Knowledge (SoK) paper on March 7, 2026, providing the first unified mathematical framework for understanding agentic Retrieval-Augmented Generation systems. The paper formalizes these increasingly autonomous AI systems as finite-horizon partially observable Markov decision processes while identifying critical reliability risks.
Mathematical Framework Models RAG as Decision-Making Systems
The paper, authored by Saroj Mishra, Suman Niroula, Umesh Yadav, Dilip Thakur, Srijan Gyawali, and Shiva Gaire, treats agentic RAG systems as sequential decision-making processes with explicit control policies and state transitions. This formalization provides theoretical foundation for analyzing systems where large language models autonomously coordinate multi-step reasoning, dynamic memory management, and iterative retrieval strategies.
The researchers developed a comprehensive taxonomy categorizing systems by their planning mechanisms, retrieval orchestration strategies, memory paradigms, and tool-invocation behaviors. The modular architectural decomposition addresses what the authors describe as "highly fragmented architectures" and "inconsistent evaluation methodologies" in current implementations.
Critical Systemic Risks Identified in Autonomous Loops
The paper identifies four severe systemic risks inherent to autonomous retrieval-generation loops:
- Compounding hallucination propagation: Errors accumulate across multiple retrieval steps, creating cascading inaccuracies
- Memory poisoning: Corrupted information persists in system memory, affecting subsequent operations
- Retrieval misalignment: Retrieved documents fail to match actual information needs
- Cascading tool-execution vulnerabilities: Errors in one tool call propagate to downstream operations
These findings challenge traditional static evaluation practices, which the researchers argue are inadequate for assessing autonomous systems that make sequential decisions.
Research Roadmap for Reliable Agentic Systems
The paper outlines research directions spanning stable adaptive retrieval, cost-aware orchestration, formal trajectory evaluation, and oversight mechanisms. The authors describe these as "doctoral-scale research directions," indicating the depth of work needed to address fundamental challenges in making agentic RAG systems reliable and controllable.
The work addresses a critical gap between rapid industrial adoption of agentic RAG architectures and systematic understanding of how these systems work, their failure modes, and reliability guarantees. The paper is available as arXiv preprint 2603.07379 in the AI, computational linguistics, cryptography, and information retrieval categories.
Key Takeaways
- First unified framework formalizes agentic RAG systems as finite-horizon partially observable Markov decision processes
- Paper identifies four critical risks: compounding hallucination, memory poisoning, retrieval misalignment, and cascading tool failures
- Comprehensive taxonomy categorizes systems by planning, retrieval orchestration, memory, and tool-invocation approaches
- Research published March 7, 2026, by six-researcher team on arXiv (ID: 2603.07379)
- Work provides roadmap for building reliable, controllable, and scalable agentic retrieval systems