This work introduces the task of hallucination span detection with input-side evidence alignment, which jointly identifies hallucinated spans and aligns output tokens with the corresponding input evidence.
Abstract
Hallucinations remain a major obstacle to the reliable use of large language models (LLMs) in conditional text generation. Existing methods primarily assess the factuality of an entire generated text, providing limited insight into which output spans are hallucinated or how they relate to the input. We introduce the task of hallucination span detection with input-side evidence alignment, which jointly identifies hallucinated spans and aligns output tokens with the corresponding input evidence. Our approach is based on the observation that faithful output tokens are predictable from the input, whereas hallucinated tokens are not. We therefore train an encoder-based model to predict masked output tokens from the input representation, using prediction confidence for hallucination detection while naturally producing alignments to the input. Experiments show that the proposed method effectively detects hallucinated spans and identifies meaningful input-side evidence. Human evaluation confirms the quality of the predicted alignments.
SeeMe is proposed, a training-free framework that introduces the concept of feature engineering from traditional machine learning into LVLMs and restructures visual tokens through a three-stage token engineering process to suppress hallucination sources while preserving informative visual evidence.
Kai Tang, Jinhao You, Bohua Zhang et al.· 2 citations
Experiments show that CORTEX substantially improves token-level hallucination detection, with each component consistently contributing to performance gains.
Kazuaki Furumai, Shuichiro Haruta, K. Matsumoto et al.· arXiv.org· 0 citations
This work proposes AURORA, a novel hallucination detection framework that shifts the focus from static representations to the weight-gradient dynamics of LLMs, and achieves strong hallucination detection performance across four model families and four benchmark datasets.
Z. Zhang, Hainan Zhang, Zhiming Zheng· arXiv.org· 0 citations
This work studies hallucination from the perspective of dynamic representation shift during generation and proposes an online projection-based intervention on intermediate hidden states to suppress the hallucination-related directions, mitigating hallucinations while preserving language quality.
Large Language Models frequently generate responses that appear fluent and coherent but may contain factually incorrect information, a phenomenon known as hallucinations. Hallucinations have emerged as a significant barrier to the effective application of Large Language Models. Some studies suggest that hallucinations can be detected by analyzing structural properties of attention maps. However, such methods primarily focus on attention structures and their spectral features, whereas uncertainty, randomness, and other statistical measures can also be leveraged to detect hallucinations. In this work, we propose a method for detecting hallucinations using spectral and statistical features extracted from attention maps. We compute eigenvalues of the Laplacian matrix from each attention map to capture spectral features and compute statistical measures for every attention head. The spectral and statistical features are used as a feature vector for hallucination detection. We conducted experiments on the HaluEval dataset across multiple language models, and the results show that jointly modeling structures and uncertainty-aware characteristics leads to more robust hallucination detection.
Rln Lokesh, Chirumamilla Mokshagna, K. Dinesh et al.· International Conference on...· 0 citations
This work studies Post Hallucination Reasoning (PHR), the stage in which hallucinated semantics enter the model's inference context and influence downstream predictions, and introduces HIVE, Hallucination Inference and Verification Engine, an evaluation infrastructure that enables controlled comparisons between faithful and hallucinated captions.
Feng He, Zhenting Wang, Qifan Wang et al.· 1 citation