Adaptive Critic-Guided Hybrid Agentic RAG for Improving Retrieval Robustness and Hallucination Resistance Through Multi-Stage Verification

MOHITH CHANDRA GUGULOTHU, Shreya Yadav, Rishu Kumar, Rajiv Misra

Keywords: etrieval-Augmented Generation, Hallucination Mitigation, Agentic AI Systems, Critic-Guided Generation, LangGraph, Adaptive Retrieval Policy.

TL;DR: A critic-guided hybrid agentic RAG framework that eliminates hallucinated responses through atomic claim verification, adaptive retrieval, and multi-stage self-corrective reasoning for trustworthy LLM generation.

Abstract:

Retrieval-Augmented Generation (RAG) systems en-
hance the factual grounding capability of large language
models (LLMs) by incorporating external knowledge during
response generation. However, conventional RAG pipelines
remain highly vulnerable to hallucinations, retrieval insta-
bility, and unreliable reasoning when handling ambiguous,
unsupported, or out-of-domain queries. This paper presents
an adaptive critic-guided hybrid agentic RAG framework de-
signed to improve hallucination resistance, retrieval robustness,
and self-correction capability in locally deployed LLM systems.
The proposed architecture integrates dense vector retrieval,
BM25 lexical retrieval, adaptive retrieval policies, sentence-
level atomic claim verification, semantic failure memory, and
web fallback mechanisms within a multi-agent Lang- Graph
orchestration pipeline. Furthermore, the framework dynami-
cally adjusts retrieval depth according to query com- plexity
and employs critic-guided verification prior to final answer
generation to improve factual reliability and safer response
behavior. Experimental evaluation demonstrates that the pro-
posed framework completely eliminated hallucinated responses
on stress-test benchmark queries, reducing the hallucination
rate from 0.34 in the baseline RAG pipeline to 0.00 by
refusing unsupported generations instead of producing con-
fident fabricated answers. The framework additionally im-
proved answer relevance from 0.71 to 0.74 and achieved a
retry effectiveness gain of +0.047 through adaptive retrieval
refinement. Although the proposed architecture introduced
approximately 3.2× higher inference latency due to sequen-
tial multi-agent verification and claim-level evidence checking
(35.92s → 116.14s), the results demonstrate that atomic claim
verification and critic-guided reasoning substantially improve
reliability, hallucination resistance, and safe handling of unsup-
ported queries. Overall, the findings highlight the effectiveness
of combining adaptive retrieval, multi-stage verification, and
semantic memory for developing more trustworthy and robust
agentic RAG systems.

Submission Number: 6