Explore the vast range of titles available.

Retrieval-augmented generation (RAG) leverages the strengths of information retrieval and generative models to enhance the handling of real-time and domain-specific knowledge. Despite its advantages, limitations within RAG components may cause hallucinations, or more precisely termed confabulations in generated outputs, driving extensive research to address these limitations and mitigate hallucinations. This review focuses on hallucination in retrieval-augmented large language models (LLMs). We first examine the causes of hallucinations from different sub-tasks in the retrieval and generation phases. Then, we provide a comprehensive overview of corresponding hallucination mitigation techniques, offering a targeted and complete framework for addressing hallucinations in retrieval-augmented LLMs. We also investigate methods to reduce the impact of hallucination through detection and correction. Finally, we discuss promising future research directions for mitigating hallucinations in retrieval-augmented LLMs.

Multimodal large language models (MLLMs) have achieved impressive performance in understanding and describing visual content, setting new state-of-the-art results on a variety of visual question answering (VQA) benchmarks. However, during decoding, these models often fail to attend to fine-grained visual details in the input image. Our analysis of intermediate attention layers reveals that MLLMs are not inherently incapable of perceiving target objects; rather, attention to visual details becomes diluted in deeper layers due to the dominance of language priors. To address this limitation, we propose a plug-and-play attention re-alignment module (ARA) that enhances suppressed visual grounding. ARA conducts a layer-wise analysis of the relative attention distribution of image-centric attention heads. It incorporates a confidence-aware layer selection mechanism based on attention peak and entropy, enabling the dynamic aggregation of attention maps from the most informative layers. These aggregated maps are subsequently leveraged to guide the generation of semantic masks, enabling the model to emphasize salient visual regions while suppressing irrelevant or noisy content. ARA can be seamlessly integrated into existing MLLMs and demonstrates consistent improvements across multiple VQA benchmarks, validating its effectiveness in enhancing visual detail sensitivity.

With the widespread application of large language models (LLMs) in natural language processing (NLP), hallucinations have become a significant impediment to their effective use of LLMs in industry applications. To address this challenge, we integrate existing hallucination detection and mitigation methods into a unified hallucination detection and mitigation framework. The framework consists of four main components: output parser, reference parser, fact verifier, and mitigator. These components collectively consolidate various hallucination detection and mitigation methods. Within this unified framework, we introduce the hierarchical semantic piece (HSP) for hallucination detection and mitigation. The HSP method extracts multi-granularity semantic pieces from both the reference material and the generated text. Sentence-level semantic pieces encapsulate global semantic information, while entity-level semantic pieces handle local semantic information. This method verifies the consistency between the generated text and the reference text at corresponding granularities, thereby enhancing the effectiveness of hallucination detection and mitigation. Experimental results show that the HSP method is very effective in detecting and mitigating hallucinations and shows lower computational resource consumption. Our method has great potential and promises for industry applications that rely on professionalism and reliability.

For clinical decision support to work, medical knowledge needs to be easy to find quickly and accurately. Retrieval-Augmented Generation (RAG) systems use big language models and document retrieval to help with diagnostic reasoning, but they could cause hallucinations and have strict privacy rules in healthcare. We tested twelve different types of RAG, such as dense, sparse, hybrid, graph-based, multimodal, self-reflective, adaptive, and security-focused pipelines, on 250 de-identified patient vignettes. We used Precision@5, Mean Reciprocal Rank, nDCG@10, hallucination rate, and latency to see how well the system worked. The best retrieval accuracy (P@5 ≥ 0.68, nDCG@10 ≥ 0.67) was achieved by a Haystack pipeline (DPR + BM25 + cross-encoder) and hybrid fusion (RRF). Self-reflective RAG, on the other hand, lowered hallucinations to 5.8%. Sparse retrieval gave the fastest response (120 ms), but it was not as accurate. We also suggest a single framework for reducing hallucinations that includes retrieval confidence thresholds, chain-of-thought verification, and outside fact-checking. Our findings emphasize pragmatic protocols for the secure implementation of RAG on premises, incorporating encryption, provenance tagging, and audit trails. Future directions encompass the incorporation of clinician feedback and the expansion of multimodal inputs to genomics and proteomics for precision medicine.

The deployment of large language models as autonomous retrieval agents over unstructured knowledge bases gives rise to a persistent structural conflict between probabilistic neural generation and deterministic physical execution. While agentic paradigms facilitate complex multi-hop retrieval, their unconstrained generative nature frequently violates strict syntactic requirements. This systemic vulnerability directly triggers execution hallucinations, such as fabricated API parameters or malformed schemas. Consequently, these syntax-driven failures force systems into redundant trial-and-error recovery loops, resulting in severe computational inflation that degrades both token efficiency and inference latency. To resolve this reliability–efficiency dilemma, this paper proposes RAG-CoT-MCP, a neuro-symbolic architecture that orthogonally decouples probabilistic cognitive planning from deterministic tool execution. By integrating the Model Context Protocol (MCP) as a strict system-level validation boundary, the framework ensures that latent reasoning trajectories manifest exclusively as syntactically valid operations. Exhaustive empirical evaluations across four disparate datasets—incorporating a multi-dimensional LLM-as-a-Judge framework, rigorous ablation studies, and granular cost tracking—validate the proposed approach. The findings demonstrate that RAG-CoT-MCP compresses network-level execution error rates from 45.2% (in unconstrained baselines) to a mere 6.0%, yielding substantial enhancements in semantic comprehensiveness and logical coherence compared to existing baselines. Counterintuitively, by proactively intercepting malformed actions and redirecting computational resources from reactive error handling to valid causal deduction, the framework drastically reduces redundant token consumption and achieves the lowest overall inference latency. Ultimately, this study establishes that deterministic execution constraints do not hinder agentic flexibility; rather, they serve as a fundamental prerequisite for deploying robust, high-speed, and cost-effective knowledge retrieval systems.

The automation of precise discipline inspection consultation requires question-answering (QA) systems that are both semantically nuanced and factually grounded. To address the limitations of keyword-based retrieval and the hallucination tendencies of generative language models in high-stakes discipline inspection domains, we propose a two-stage Retrieval-Augmented Generation (RAG) framework designed for Chinese discipline inspection text. Our approach synergizes token-level late interaction and cross-encoder reranking to achieve high-precision evidence retrieval. First, we employ ColBERTv2 to perform efficient, fine-grained semantic matching between queries and lengthy discipline inspection documents. Subsequently, we refine the initial candidate set using a computationally focused cross-encoder, which performs deep pairwise relevance scoring on a shortlist of passages. This retrieved evidence strictly conditions the answer generation process of a large language model (DeepSeek-chat). Through rigorous evaluation on a curated corpus of real Chinese discipline inspection documents and expert-annotated queries, we demonstrate that our pipeline significantly outperforms strong baselines—including BM25, single-stage dense retrieval (BGE), and a simplified ColBERT variant—in both retrieval metrics (Recall@k, Precision@k) and answer faithfulness. Our work provides a robust, reproducible blueprint for building reliable, evidence-based discipline inspection AI systems, highlighting the critical role of hierarchical retrieval in mitigating hallucinations for domain-specific QA.

Biodiversity knowledge is fundamental to conservation planning and sustainable environmental decision-making; however, general-purpose Large Language Models (LLMs) frequently produce hallucinations when responding to biodiversity-related queries. To address this challenge, we propose BioChat, a domain-specific question-answering system that integrates a Retrieval-Augmented Generation (RAG) framework with a Re-Ranker–based retrieval and routing mechanism. The system is built upon a verified biodiversity dataset curated by the National Institute of Biological Resources (NIBR), comprising 25,593 species and approximately 970,000 structured data points. We systematically evaluate the effects of embedding selection, routing strategy, and generative model choice on factual accuracy and hallucination mitigation. Experimental results show that the proposed Re-Ranker-based routing strategy significantly improves system reliability, increasing factual accuracy from 47.9% to 71.3% and reducing hallucination rate from 34.0% to 24.4% compared with Naive RAG baseline. Among the evaluated LLMs, Qwen2-7B-Instruct achieves the highest factual accuracy, while Gemma-2-9B-Instruct demonstrates superior hallucination control. By delivering transparent, verifiable, and context-grounded biodiversity information, BioChat supports environmental education, citizen science, and evidence-based conservation policy development. This work demonstrates how trustworthy AI systems can serve as sustainability-enabling infrastructure, facilitating reliable access to biodiversity knowledge for long-term ecological conservation and informed public decision-making.

IntroductionMultimodal large language models (MLLMs) have exhibited significant potential for fisheries analysis. However, their inherent hallucination issues in species identification, ecological activity interpretation, environmental assessment, and biomass estimation severely restrict their reliability and practical application in real‑world fisheries management.MethodsThis study proposes FMRAG, a fisheries‑oriented multimodal retrieval‑augmented generation framework to enhance factual grounding and domain adaptability. The framework retrieves visually similar fishery images and corresponding textual records within a unified vision‑language embedding space, and integrates retrieved multimodal evidence with query inputs via a cross‑modal fusion mechanism to exploit fine‑grained visual features and domain‑specific textual knowledge. Three fisheries‑tailored fine‑tuning tasks are further introduced: image‑text association learning, visual concentration learning, and retrieval‑augmented reasoning learning, to strengthen multi‑image reasoning and multimodal alignment.ResultsExperimental results on species identification and biomass estimation demonstrate that FMRAG consistently outperforms baseline MLLMs and text‑only RAG methods, effectively reducing hallucinations and improving predictive accuracy. The proposed framework also shows superior performance in rare‑species recognition, temporal stability, and confidence calibration, and can be successfully transferred to models originally trained on single‑image inputs.DiscussionFMRAG provides an effective and practical solution for constructing trustworthy multimodal intelligence systems, supporting reliable and robust applications in fisheries monitoring and management.

When LLMs are applied in the veterinary field, they often produce serious hallucinations and logical restrictions, especially in the accurate diagnosis of bovine disease, where accuracy is crucial. To meet this challenge, this paper proposes GraphRAG-Vet, a Knowledge Graph Retrieval-Augmented Generation framework specifically designed for the dairy industry. First, we constructed a domain knowledge map comprising 2500 elements and 3000 relationships, covering high-frequency diseases in cows such as mastitis and ketosis. Second, the semantic-to-password parsing module is designed to retrieve disease symptom subgraphs from the Neo4j database accurately. Finally, the hard constraint injection mechanism is introduced to force LLMs to generate diagnoses strictly in accordance with the retrieved graph context, thereby implementing the “refuse to answer” function for foreign queries. The experimental results showed that GraphRAG-Vet achieved 100% accuracy in diagnosing core infectious diseases and had an almost-zero hallucination rate compared with baseline LLMs. This study provides a reliable, low-resource solution for automated veterinary consultation.

Recent image captioning methods based on pre-trained vision–language models can generate fluent and coherent descriptions, yet they still struggle in open-domain scenes that contain long-tail concepts, uncommon object combinations, and ambiguous visual evidence. Two limitations are especially important. First, the knowledge needed to recognize and name rare or domain-specific entities is only weakly represented in model parameters, causing captions to be generic, incomplete, or biased toward frequent concepts. Second, token generation is typically grounded mainly by local visual matching, making it sensitive to clutter, occlusion, and visually similar distractors, and therefore prone to attribute errors, relation confusion, and object hallucination. To address these issues, we propose R2G (retrieval- and grounding-guided captioning), a lightweight plug-in framework for frozen image captioning backbones. R2G consists of two complementary components. The first, retrieval-guided visual prompting, retrieves image-relevant concepts from an external visual concept memory, converts them into a continuous prompt representation, and injects this representation into selected layers of the visual encoder, so that external semantic information can influence visual feature formation before decoding begins. The second, global–local semantic grounding, derives a global semantic prior from an auxiliary vision–language encoder and adaptively fuses it with token-level local visual evidence through a decoder-state-dependent gating mechanism, thereby improving semantic stability while preserving fine-grained visual support. The resulting framework is lightweight, compatible with frozen pre-trained backbones, and designed to improve both concept coverage and semantic faithfulness. Experimental results on MS-COCO and NoCaps show that R2G consistently improves caption quality over the baseline and yields particularly clear gains in open-domain and out-of-domain settings.