Adaptive Retrieval-Augmented Generation Agents with Autonomous Evidence Verification

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Large language models (LLMs) can generate fluent and informative text, yet their internal parametric knowledge is often incomplete, outdated, or incorrectly recalled, leading to unsupported or factually inconsistent claims. Retrieval-Augmented Generation (RAG) mitigates this problem by conditioning generation on externally retrieved evidence, but conventional RAG pipelines retrieve a fixed number of documents regardless of query difficulty and rarely verify whether the retrieved evidence actually supports the generated answer. This paper proposes an Adaptive Retrieval-Augmented Generation agent (AR-Verify) that treats retrieval and verification as iterative, evidence-driven decisions rather than static pipeline stages. The agent decomposes complex queries into sub-claims, dynamically selects retrieval strategies based on estimated evidence sufficiency, and invokes a verification module that checks generated claims against retrieved passages prior to finalizing a response. When evidence is judged insufficient, contradictory, or of low quality, the agent autonomously reformulates queries, expands retrieval to alternative sources, or requests additional targeted evidence for specific claims. We describe the agent architecture, the evidence sufficiency and claim-verification formulations, and an evaluation protocol comparing the proposed approach against fixed-pipeline RAG baselines on retrieval accuracy, answer correctness, faithfulness, evidence coverage, hallucination rate, and computational cost. Representative target results, reported here to illustrate the intended evaluation depth for an eventual full implementation, indicate that adaptive retrieval and verification can reduce unsupported claims relative to fixed-pipeline RAG at a moderate increase in latency and retrieval cost. We discuss architectural trade-offs, limitations of self-verification, and directions for reducing the added computational overhead of iterative evidence checking.