Tag
LongNovel introduces a multi-scale benchmark for evaluating hallucinations in long-context novel summarization, built from Chinese and English novels to improve LLM assessment.
This paper introduces InnerExpert, a method that leverages Mixture-of-Experts signals for per-token hallucination detection in LLMs, achieving superior performance over existing techniques.
This paper investigates how Large Language Models compress conceptual distances in long contexts into short navigable paths, demonstrating the 'Six Degrees of Separation' phenomenon in their latent spaces, and applies this framework to detect hallucinations in Retrieval-Augmented Generation systems.
BEAR-Bench introduces a bilingual English-and-Russian benchmark for evaluating multimodal models' reasoning on text-rich professional documents, assessing 16 models and highlighting performance gaps.
This paper introduces a task for hallucination span detection in LLMs by aligning output tokens with input evidence, proposing an encoder-based model that uses prediction confidence to detect hallucinations without manual alignments.
This paper proposes DeMTS, a framework that treats denoising trajectories in diffusion language models as multivariate time series to improve hallucination detection, outperforming existing methods with strong robustness and efficiency.
This paper identifies and characterizes Decomposition-Induced Context-Memory Conflict (DI-CC), a failure mode in decompose-then-verify pipelines where decomposition substitutes the model's parametric beliefs for the source text. The authors show it is mechanistically related to classical context-memory conflict, that SelfCheckGPT fails to detect it, and that context-aware decoding suppresses it but introduces severe parsing failures.
This paper introduces the Latent Critic, a lightweight LoRA adapter that detects hallucinated agent actions in real time by restructuring the transformer's residual stream into localized natural-language feedback, achieving 0.966 AUROC and enabling self-correction.
UniProbe is a lightweight detector that uses graph-based multi-structural internal representations to identify hallucinated tokens in large vision-language models, reducing object hallucinations by up to 55% with minimal latency increase.
This paper introduces Prompt Embedding Probes (PEP), a parameter-efficient extension of linear probes that uses learnable prompt embeddings on hidden states to detect hallucinations in frozen LLMs. Evaluations on TriviaQA, GSM8K, and MedQA with Qwen3 models show improvements over standard linear probes, including in pre-generation and cross-model settings.
This paper presents HallDetect, a lightweight and reference-free framework for hallucination detection that decomposes generated content into atomic claims and verifies them with a compact entailment model. It outperforms comparably resourced baselines on multiple benchmarks and provides a claim-to-span audit trail.
This paper introduces TriQua, a framework for LLM factuality evaluation that adaptively represents facts as triples or hyperrelational facts with contextual qualifiers, along with TriQuaScore for fine-grained factuality scoring. It demonstrates strong alignment with human annotations and improved evidence-based verification over existing methods.
A pre-registered study proposes a first-token hallucination detection method using internal model signals across 10 models, finding no universal detector but a universal above-chance floor, with public code and verification scripts.
Introduces ReDe, a framework that denoises reasoning traces by filtering irrelevant and repetitive steps to improve hallucination detection in large reasoning models, achieving up to 87.32 AUROC on TruthfulQA.
The paper proposes a neurosymbolic framework that lifts LLM outputs into typed compositional derivations using Combinatory Categorial Grammar, enabling structural checking and early hallucination detection.
This paper proposes Logical Graph Uncertainty (LGU), a framework that models implication and incompatibility among answers to improve uncertainty estimation in LLMs, outperforming semantic entropy baselines by up to 7.1% AUROC and 3.5% AUARC across benchmarks.
This paper proposes diversity-oriented fine-tuning strategies to improve uncertainty-based hallucination detection in LLMs by encouraging varied generations, making hallucinations more detectable via semantic entropy.
ASK-NN is an asymmetric nearest-neighbor test for detecting distribution drifts between reference and query samples, with applications to LLM hallucination detection and artificial-text detection. It is computationally efficient, has theoretical guarantees, and performs competitively against baselines on synthetic and real-world benchmarks.
An open-source tool for training LLMs on consumer hardware, featuring real-time neural visualization for hallucination detection and model introspection, currently supporting small-scale models.
This paper introduces diversion decoding, a novel method for detecting hallucinations in large language models by actively challenging model responses during the decoding phase to extract features for training an uncertainty heuristic, achieving superior performance with lower computational complexity.