hallucination-detection

This paper proposes automatic layer selection for hallucination detection in LLMs and introduces First Effective Peak of Intrinsic Dimension (FEPoID), a training-free criterion that consistently identifies optimal intermediate layers, outperforming existing heuristics.

This paper introduces Graph Alignment Topology as an inductive bias for grounding detection, using a graph neural network to model alignment structure between reference information and LLM outputs. The method achieves state-of-the-art results on multiple hallucination and question-answering datasets, outperforming GPT-4o.

This paper reveals that much of the reported progress in LLM hallucination detection is due to benchmark construction artifacts, where ground-truth answers are embedded in prompts, allowing a simple text-similarity baseline to achieve near-perfect scores. Through a large-scale controlled evaluation, the authors show that most methods perform near chance under proper controls, except for supervised probes on upper-layer hidden states such as SAPLMA and their proposed DRIFT.

This paper presents the first systematic study of uncertainty quantification (UQ) for Large Language Diffusion Models (LLDMs), proposing lightweight zero-shot uncertainty signals derived from the iterative denoising process and showing that LLDMs can achieve both fast inference and reliable hallucination detection with up to 100x lower computational overhead compared to sampling-based baselines.

Academic Research Skills is the first installable Claude Code workflow that packages a multi-agent pipeline to detect and prevent hallucinated citations in academic papers, addressing a problem where 146,932 hallucinated citations were counted in 2025 preprints.

This paper analyzes hallucination detection in LLMs, proposing a max-pooling approach that improves efficiency by eliminating costly semantic consistency computations while maintaining competitive performance.

This paper introduces CiteTracer, a multi-agent framework for detecting citation hallucinations in LLM-generated scientific writing, achieving high accuracy on synthetic and real-world benchmarks.

This paper investigates whether standard benchmarks underestimate LLM performance by re-evaluating hallucination detection datasets using an LLM-first, human-adjudicated assessment method. The study finds that incorporating LLM reasoning into the adjudication process improves agreement and suggests that model-assisted re-evaluation yields more reliable benchmarks for ambiguity-prone tasks.

This paper introduces a controlled-invariance methodology and two oracle tests (Force and Remove) to determine if LLM hallucination detectors rely on reasoning traces or final answer artifacts. It proposes TRACT, a lightweight scorer using lexical features, which demonstrates robust performance independent of answer-level cues.

This paper introduces a proxy-analyzer framework that detects hallucinations in large language models by analyzing internal activations of small, open-weight models rather than the generator itself. The method achieves superior performance on benchmarks like RAGTruth compared to existing methods like ReDeEP, demonstrating that model size is less critical than the analysis approach.

This paper presents PCNet, a probabilistic circuit trained as a tractable density estimator on LLM residual streams to detect hallucinations as geometric anomalies. It also introduces PC-LDCD, a dynamic correction method that only intervenes on hallucinated tokens, achieving near-perfect detection and reduced corruption rates.

This paper introduces a method for detecting hallucinations in large language models by leveraging the confidence of the first generated token, requiring only a single decode step.

Researchers introduce SHADE, a hybrid estimator that combines Good-Turing coverage with graph-spectral cues to quantify semantic uncertainty and detect LLM hallucinations when only a few black-box samples are available.

Researchers from Beihang University and other institutions propose HalluSAE, a framework using sparse autoencoders and phase transition theory to detect hallucinations in LLMs by modeling generation as trajectories through a potential energy landscape and identifying critical transition zones where factual errors occur.

TPA proposes a novel method for detecting hallucinations in RAG systems by attributing next-token probabilities to seven distinct sources (Query, RAG Context, Past Token, Self Token, FFN, Final LayerNorm, Initial Embedding) and aggregating by Part-of-Speech tags. The approach achieves state-of-the-art performance across five LLMs including Llama2, Llama3, Mistral, and Qwen.

This paper introduces FRANQ, a method for detecting hallucinations in Retrieval-Augmented Generation (RAG) systems by applying distinct uncertainty quantification techniques to distinguish between factuality and faithfulness to retrieved context. The authors construct a new dataset annotated for both factuality and faithfulness, and demonstrate that FRANQ outperforms existing approaches in detecting factual errors across multiple datasets and LLMs.

RAGognizer introduces a hallucination-aware fine-tuning approach that integrates a lightweight detection head into LLMs for joint optimization of language modeling and hallucination detection in RAG systems. The paper presents RAGognize, a dataset of naturally occurring closed-domain hallucinations with token-level annotations, and demonstrates state-of-the-art hallucination detection while reducing hallucination rates without degrading language quality.

This paper introduces SIVR (Sequential Internal Variance Representation), a supervised framework for detecting hallucinations in LLMs by analyzing token-wise and layer-wise variance patterns in hidden states without relying on strict architectural assumptions. The method aggregates full sequence variance features to learn temporal patterns of factual errors and demonstrates improved generalization with smaller training sets.

OpenAI introduces SimpleQA, a new factuality benchmark dataset with 4,326 short fact-seeking questions designed to evaluate frontier language models on their ability to provide accurate answers without hallucination. The dataset achieves high quality through dual independent annotation, rigorous criteria, and achieves only ~3% estimated error rate, with GPT-4o scoring less than 40%.