retrieval-augmented generation

Hallucination remains a critical reliability challenge in the practical deployment of large language models (LLMs). Method: This work introduces, for the first time, a dual-dimensional taxonomy of hallucinations—distinguishing *knowledge-based* and *logic-based* types—and proposes a unified framework integrating retrieval-augmented generation (RAG), chain-of-thought (CoT) reinforcement, and agent-based system orchestration to systematically mitigate them. We analyze the intrinsic mechanisms by which each component suppresses distinct hallucination categories. Contribution/Results: Through rigorous empirical evaluation on standardized benchmarks, we systematically characterize the suppression pathways of each technique across hallucination types. The study delivers a reusable, modular paradigm for enhancing LLM reliability and a standardized evaluation framework. Our approach significantly improves both factual accuracy and operational feasibility—bridging the gap between theoretical robustness and real-world deployment.

To address core limitations of large language models (LLMs)—including hallucination, knowledge obsolescence, and poor domain adaptability—this work systematically advances the Retrieval-Augmented Structured (RAS) generation paradigm. We propose a multi-granularity knowledge acquisition mechanism integrating sparse, dense, and hybrid retrieval, coupled with text structuralization, taxonomy construction, knowledge embedding, and prompt-driven reasoning to enable efficient external knowledge retrieval, semantic alignment, and controllable integration. Crucially, we deeply embed structured modeling into the augmentation pipeline, enhancing factual accuracy, temporal freshness, and domain-specific competence of generated outputs. Our contributions include: (1) a unified methodological framework for RAS generation; (2) principled pathways toward multimodal, cross-lingual, and interactive augmented generation; and (3) empirically validated improvements in reliability and specialization across diverse domains. This work establishes foundational design principles and future research directions for next-generation RAS systems.

The KG-RAG field suffers from a lack of systematic understanding and reproducible benchmarking. Method: We conduct a comprehensive evaluation across six KG-RAG methods, seven datasets, and seventeen large language models (LLMs), establishing the first applicability criterion system for KG-RAG. Through knowledge graph (KG) embedding optimization, RAG pipeline reconstruction, and multidimensional ablation studies, we propose a “need-based activation” paradigm—replacing default KG integration with selective, task-aware utilization. Contribution/Results: Our analysis reveals a strong coupling between KG components and specific task scenarios, identifying three high-benefit settings: fact-intensive question answering, multi-hop reasoning, and low-resource domains. Empirical results demonstrate that optimal configuration reduces hallucination rates by 32% and improves answer accuracy by up to 27%, providing principled guidance for effective KG-RAG deployment.

This study addresses the theoretical limitations of traditional dense retrieval and the vulnerability of generative retrieval under document identifier ambiguity. For the first time, it systematically evaluates the potential of generative retrieval on the synthetic dataset LIMIT, employing SEAL and MINDER models with BM25 and dense retrieval as baselines. Results show that on the original LIMIT dataset, generative approaches achieve R@2 scores of 0.92–0.99, substantially outperforming dense retrieval (<0.03) and BM25 (0.86). However, when hard negatives are introduced, performance sharply drops to 0.51, revealing a critical bottleneck: the decoding mechanism struggles to generate unique identifiers reliably. This work not only confirms the superiority of generative retrieval under ideal conditions but also, through error analysis, identifies identifier ambiguity as a key challenge limiting its robustness.

Can generative retrieval (GR) overcome the representation and optimization bottlenecks inherent in dense retrieval (DR)? This work systematically investigates the fundamental differences between GR and DR—in learning objectives, representational capacity, and optimization mechanisms—from both theoretical and empirical perspectives. We show that GR employs globally normalized maximum-likelihood optimization, directly modeling corpus distributions and relevance relationships, thereby avoiding the optimization drift commonly observed in DR; moreover, its representational capacity scales with model size and is not constrained by low-rank embedding assumptions. Experiments on Natural Questions and MS MARCO demonstrate GR’s superior scalability and theoretical advantages, though its current performance does not yet consistently surpass state-of-the-art DR methods. The study identifies critical influences of negative sampling strategies, model scale, and bilinear similarity design, and proposes concrete directions for advancing GR toward practical deployment.

In the era of pervasive generative AI—particularly large language models—a critical gap has emerged between the speed of information generation and the pace of factual verification, enabling rapid dissemination of misinformation. Method: We propose the RGB (Generation–Indexing–Broadcasting) quantitative framework, modeling the information lifecycle as a stochastic process to systematically characterize the dynamics across generation, indexing, and broadcasting stages. Leveraging stochastic process theory, empirical time-series analysis, and large-scale data mining from Stack Exchange, we quantify verification delays in Retrieval-Augmented Generation (RAG) systems, especially for emerging topics. Contribution/Results: We demonstrate that high-quality, factually grounded answers require substantial human effort and incur inherent temporal latency; critically, current AI generation rates systematically outpace human verification capacity, escalating misinformation risk. This work establishes the first theoretical foundation and quantitative toolkit for assessing RAG reliability and enabling trustworthy retrieval-augmented inference.

Generative retrieval—where large language models (LLMs) autoregressively generate document identifiers—lacks a clear understanding of how model size, training data volume, and inference compute jointly scale. Method: We conduct the first systematic study of this triadic scaling relationship, introducing a continuous evaluation metric that integrates contrastive entropy and generative loss to enable robust, architecture-agnostic comparisons. Using a unified framework across LLaMA (decoder-only) and T5 (encoder-decoder), we combine n-gram analysis with large-scale ablation experiments. Contributions/Results: All three resources—model scale, data volume, and inference compute—exhibit strong positive correlations with retrieval performance. LLaMA consistently outperforms T5 across multiple configurations. Crucially, n-gram modeling adheres strictly to power-law scaling behavior, offering a novel, interpretable foundation for generative retrieval. Our findings establish principled guidelines for resource-aware model design and deployment in generative retrieval systems.

This work addresses the challenge of faithfulness hallucinations in retrieval-augmented generation (RAG) systems, which often arise from the opaque internal reasoning of large language models and evade detection by existing methods due to their inability to capture fine-grained semantic dependencies. To overcome this limitation, the authors propose a novel semantic-level internal reasoning graph approach that elevates reasoning graphs from the token level to the semantic level. By leveraging an enhanced layer-wise relevance propagation algorithm, the method explicitly models the model’s semantic reasoning pathways. Combined with a compact pretrained language model and a dynamic thresholding mechanism, this framework enables efficient and precise hallucination detection. Evaluated on the RAGTruth and Dolly-15k benchmarks, the proposed approach significantly outperforms state-of-the-art methods, demonstrating improved accuracy and robustness in identifying faithfulness hallucinations.

This work addresses the limitations of existing retrieval-augmented generation (RAG) approaches, which rely on vector similarity retrieval and are thus susceptible to semantic noise, often failing to satisfy the multi-constraint requirements of factual queries. To overcome this, the authors propose Structure-Guided Retrieval-Augmented Generation (SG-RAG), a novel framework that formalizes the “Exact Retrieval Problem” (ERP) and reframes retrieval as an embedding-based subgraph matching task. By explicitly integrating the topological structure of knowledge graphs with semantic embeddings, SG-RAG guides large language models to generate responses that strictly adhere to all query constraints. Experiments on the newly constructed large-scale ERQA dataset demonstrate that SG-RAG substantially outperforms strong baselines, achieving absolute gains of 20.68–50.88 points on key metrics while maintaining reasonable computational overhead.

Large language models are prone to hallucinations in complex mathematical reasoning due to reliance on static internal knowledge. This work proposes an adaptive retrieval-augmented architecture that enables the model to actively decide, during inference, whether to consult an external knowledge base, treating retrieval as a dynamic form of in-context learning. The study reveals that the model’s decision not to retrieve serves as a strong metacognitive signal of high performance, with retrieval providing significant benefits only in specific scenarios—such as when citing critical theorems. Evaluated on GSM8K and MATH-500 benchmarks, the approach combined with chain-of-thought (CoT) reasoning outperforms standard CoT even when no retrieval occurs, and dynamically adjusts retrieval frequency based on problem difficulty. These findings underscore the crucial role of self-assessment and selective retrieval in enhancing reasoning robustness.

This work addresses the limited transparency and explainability of traditional Retrieval-Augmented Generation (RAG) systems, whose fully neural architectures hinder interpretability and debugging in high-stakes applications. To overcome this, the authors propose a novel paradigm that integrates symbolic reasoning with neural retrieval by deeply embedding knowledge graphs into the RAG pipeline. The approach introduces three key mechanisms: Knowledge-Modulated Alignment Retrieval (MAR), Knowledge Graph Path–enhanced Querying (KG-Path RAG), and Process Knowledge–infused Re-ranking (Process Knowledge-infused RAG). Evaluated on a mental health risk assessment task, the proposed framework significantly enhances both the transparency of retrieved evidence and overall system performance, demonstrating the benefits of synergistically combining structured symbolic knowledge with neural retrieval.

This work addresses key limitations of traditional Retrieval-Augmented Generation (RAG) approaches, which rely on unstructured text passages and often suffer from contextual redundancy, weak semantic alignment, and fragmented reasoning chains—compromising generation quality and inflating token consumption. To overcome these issues, the authors propose Tri-RAG, a novel framework that automatically transforms external knowledge into structured “condition–proof–conclusion” triplets aligned with the reasoning process. Leveraging conditions as semantic anchors enables precise retrieval, while a lightweight, prompt-driven triplet representation and semantic matching mechanism explicitly model logical relationships among knowledge elements—all without fine-tuning the underlying large language model. Experimental results demonstrate that Tri-RAG significantly improves retrieval accuracy and reasoning efficiency across multiple benchmarks, yielding more stable generations while substantially reducing context token usage.