Organizations: Computational Unit, Z.H. College of Engineering & Technology, Aligarh Muslim University, India · Interdisciplinary Center for Artificial Intelligence, Aligarh Muslim University, India · Department of Computer Engineering, Aligarh Muslim University, India
Large Language Models (LLMs) have achieved unprecedented fluency but remain susceptible to "hallucinations" - the generation of factually incorrect or ungrounded content. This limitation is particularly critical in high-stakes domains where reliability is paramount. We propose a domain-grounded tiered retrieval and verification architecture designed to systematically intercept factual inaccuracies by shifting LLMs from stochastic pattern-matchers to verified truth-seekers. The proposed framework utilizes a four-phase, self-regulating pipeline implemented via LangGraph: (I) Intrinsic Verification with Early-Exit logic to optimize compute, (II) Adaptive Search Routing utilizing a Domain Detector to target subject-specific archives, (III) Refined Context Filtering (RCF) to eliminate non-essential or distracting information, and (IV) Extrinsic Regeneration followed by atomic claim-level verification. The system was evaluated across 650 queries from five diverse benchmarks: TimeQA v2, FreshQA v2, HaluEval General, MMLU Global Facts, and TruthfulQA. Empirical results demonstrate that the pipeline consistently outperforms zero-shot baselines across all environments. Win rates peaked at 83.7% in TimeQA v2 and 78.0% in MMLU Global Facts, confirming high efficacy in domains requiring granular temporal and numerical precision. Groundedness scores remained robustly stable between 78.8% and 86.4% across factual-answer rows. While the architecture provides a robust fail-safe for misinformation, a persistent failure mode of "False-Premise Overclaiming" was identified. These findings provide a detailed empirical characterization of multi-stage RAG behavior and suggest that future work should prioritize pre-retrieval "answerability" nodes to further bridge the reliability gap in conversational AI.
Figures & tables
Figure 1: Consequences of LLM Hallucinations
Figure 2: Intrinsic vs Extrinsic Hallucination
Type
Nature of Error
Example
Primary Cause
Factual
Incorrect statements contradicting truth
“Albert Einstein won two Nobel Prizes in Physics.”
Inaccurate memory / data bias
Numerical
Wrong figures or quantities
“The Earth’s atmosphere contains 40% oxygen.”
Lack of numerical reasoning
Logical
Invalid reasoning chain
“All doctors are scientists; Li is a scientist; therefore, Li is a doctor.”
Weak inferential ability
Commonsense
Violates basic knowledge
“The sun rises in the west and sets in the east.”
Lack of world grounding
Temporal
Outdated or mixed-up timelines
“Barack Obama is the current President of the United States.”
Static knowledge
Attributional
Wrong author or source
“The theory of relativity was proposed by Isaac Newton.”
Entity confusion
Table 1: Hallucination Types based on Factual Deviation
Figure 3: Proposed Framework
Benchmark
N
Domain
Query Type
Key Challenges
TimeQA v2
150
Temporal biographical
‘Who held role X between dates A–B?’
Sparse time-window documentation; correct refusal on unknowable queries
FreshQA v2
150
Current events + false premises
Factual + denial queries
39% of queries embed false or unresolved premises; requires refusal, not inference
HaluEval General
150
Open domain (mixed)
General knowledge, science, coding
Broad coverage; retrieval advantage lower vs. parametrically-strong baseline
MMLU Global Facts
50
Global statistics
Precise numerical recall
Exact figures unavailable in small-model memory; short answers
TruthfulQA
150
Myths / misconceptions
Myth-busting; denial of false premises
Must resist confabulation; false-premise resistance essential
Table 2: Benchmark summary
Benchmark
N
Proposed Wins
Tie
Baseline Wins
Win Rate
Hallucination
Groundedness
TimeQA v2
86*
72
10
4
83.7%
13.6%
86.4%
MMLU Global Facts
50
39
8
3
78.0%
33.1%
66.9%
FreshQA v2
150
97
37
16
64.7%
3.5%†
19.2%†
TruthfulQA
150
82
56
12
54.7%
15.1%
84.9%
HaluEval General
150
75
45
30
50.0%
21.2%
78.8%
Combined (650)
586
365
220**
65
62.3%
−
−
Table 3: Primary results across all five benchmarks
Figure 4: Benchmark Win Rate Comparison
Failure Mode
Frequency
Primary Benchmarks
Primary Root Cause
Open-Domain Parametric Competition
∼ 30
HaluEval
High-quality training data in the baseline model matches the quality of retrieved web sources for common knowledge.
False-Premise Overclaiming
∼ 14
TimeQA, FreshQA, TruthfulQA
Lack of a pre-retrieval answerability check; the system answers the premise instead of correcting it.
Vagueness (Verbose vs. Concise)
∼ 12
FreshQA, TruthfulQA
The generator produces long, hedged responses that the judge finds less useful than the baseline’s concise denial.
Retrieval Distraction
∼ 9
TruthfulQA, HaluEval
Tangential information in a retrieved document leads the generator away from the direct answer.
Numerical precision / Data mismatch
∼ 5
MMLU, FreshQA
The retrieved source utilizes different aggregation methods, timeframes, or units than the generated claim.
Structured Data Extraction Errors
∼ 4
FreshQA, MMLU
Small-model limitations (Llama 3.1 8B) leading to misreading of markdown tables or extracting wrong values from documents.
Table 4: Summary of Failure Modes
Figure 5: Realibility Heatmap (Benchmarks vs Error Types)
Retrieval-Augmented Generation (RAG) significantly improves the factuality of Large Language Models (LLMs), yet standard pipelines often lack mechanisms to verify inter- mediate reasoning, leaving them vulnerable to hallucinations in high-stakes domains. To address this, we propose a domain-specific RAG framework that integrates explicit rea- soning and faithfulness verification. Our architecture augments standard retrieval with neural query rewriting, BGE-based cross-encoder reranking, and a rationale generation module that grounds sub-claims in specific evidence spans. We further introduce an eight-category verification taxonomy that enables fine-grained assessment of rationale faithfulness, distinguishing between explicit and implicit support patterns to facilitate structured error diagnosis. We evaluate this framework on the BioASQ and PubMedQA benchmarks, specifically analyzing the impact of dynamic in-context learning and rerank- ing under constrained token budgets. Experiments demonstrate that explicit rationale generation improves accuracy over vanilla RAG baselines, while dynamic demonstration selection combined with robust reranking yields further gains in few-shot settings. Using Llama-3-8B-Instruct, our approach achieves 89.1% on BioASQ-Y/N and 73.0% on Pub- MedQA, competitive with systems using significantly larger models. Additionally, we perform a pilot study combining human expert assessment with LLM-based verification to explore how explicit rationale generation improves system transparency and enables more detailed diagnosis of retrieval failures in biomedical question answering.
Eeham Khan, Luis Rodriguez, Marc Queudot
Concordia University, Montréal, QC, Canada · Centre de recherche informatique de Montréal (CRIM), Montréal, QC, Canada
Hallucination remains a critical bottleneck for large language models (LLMs), undermining their reliability in real-world applications, especially in Retrieval-Augmented Generation (RAG) systems. While existing hallucination detection methods employ LLM-as-a-judge to verify LLM outputs against retrieved evidence, they suffer from inherent confirmation bias, where the verifier inadvertently reproduces the errors of the original generation. To address this, we introduce Multi-Agent Reinforced Self-Check for Hallucination (MARCH), a framework that enforces rigorous factual alignment by leveraging deliberate information asymmetry. MARCH orchestrates a collaborative pipeline of three specialized agents: a Solver, a Proposer, and a Checker. The Solver generates an initial RAG response, which the Proposer decomposes into claim-level verifiable atomic propositions. Crucially, the Checker validates these propositions against retrieved evidence in isolation, deprived of the Solver's original output. This well-crafted information asymmetry scheme breaks the cycle of self-confirmation bias. By training this pipeline with multi-agent reinforcement learning (MARL), we enable the agents to co-evolve and optimize factual adherence. Extensive experiments across hallucination benchmarks demonstrate that MARCH substantially reduces hallucination rates. Notably, an 8B-parameter LLM equipped with MARCH achieves performance competitive with powerful closed-source models. MARCH paves a scalable path for factual self-improvement of LLMs through co-evolution. The code is at https://github.com/Qwen-Applications/MARCH.
Zhuo Li, Yupeng Zhang, Pengyu Cheng +8
Qwen Large Model Application Team, Alibaba · The Chinese University of Hong Kong, Shenzhen
Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse natural language processing tasks, yet they remain susceptible to hallucinations -- generating content that is factually incorrect, unfaithful to provided context, or misaligned with user instructions. We present HalluScan, a comprehensive benchmark framework that systematically evaluates hallucination detection and mitigation across 72 configurations spanning 6 detection methods, 4 open-weight model families, and 3 diverse domains. We introduce three key contributions: (1) HalluScore, a novel composite metric that achieves a Pearson correlation of r = 0.41 with human expert judgments; (2) Adaptive Detection Routing (ADR), an intelligent routing algorithm achieving 2.0x cost reduction with only 0.1% AUROC degradation; and (3) systematic error cascade decomposition revealing substantial variation in hallucination error types across domains. Our experiments reveal that NLI Verification achieves the highest overall AUROC of 0.88, while RAV achieves the second-highest AUROC of 0.66.