Text Corpora as Concept Fields: Black-Box Hallucination and Novelty Measurement

Problem
This paper addresses the gap in the literature regarding the measurement of black-box hallucination and novelty in text corpora, particularly in the context of large language models (LLMs). The authors propose a novel framework, termed Concept Field, which quantifies the local drift of sentence embeddings to assess the groundedness of generated text. This work is presented as a preprint and has not yet undergone peer review.

Method
The core technical contribution is the introduction of the Concept Field, which is defined as a local drift field in sentence-embedding space. The authors compute the pointwise uncertainty of sentence transitions by evaluating the mean absolute z-distance ($ζ$) between the observed delta (the difference between consecutive sentence embeddings) and the local Gaussian estimate of the field. This scoring mechanism is designed to be black-box, meaning it does not rely on the internal workings of the model generating the sentences. The authors also introduce a Vector Sequence Database (VSDB) that efficiently stores embeddings along with metadata such as sequence position and next-delta information, facilitating the computation of the Concept Field. The evaluation is conducted on two large-scale datasets: the U.S. Code of Federal Regulations for hallucination detection and Project Gutenberg for novelty detection. The authors utilize controlled rewrites generated by LLMs to assess the performance of their method.

Results
The Concept Field framework demonstrates strong selective classification performance in distinguishing between grounded, ungrounded, and unsure sentences, achieving comparable results to retrieval-centric baselines while maintaining similar coverage-risk behavior across both datasets. The authors report that their method outperforms traditional approaches in terms of interpretability and probabilistic interpretation, although specific numerical performance metrics (e.g., accuracy, F1 scores) are not disclosed in the abstract. The qualitative analysis of divergence and curl within the Concept Field reveals meaningful semantic patterns, which the authors suggest can serve as hypothesis-generating insights.

Limitations
The authors acknowledge that their approach is primarily qualitative in nature, particularly in the analysis of divergence and curl, which may not yield quantitative results suitable for rigorous evaluation. Additionally, the reliance on LLM-generated rewrites may introduce biases inherent to the models used, potentially affecting the generalizability of the findings. The paper does not address the computational efficiency of the VSDB in large-scale applications or the scalability of the Concept Field framework across diverse text corpora.

Why it matters
The introduction of Concept Fields provides a novel, interpretable mechanism for assessing groundedness and novelty in text generation, which is crucial for improving the reliability of LLMs in practical applications. By offering a probabilistic framework that is corpus-attributable, this work opens avenues for further research into the interpretability of LLM outputs and the development of more robust detection mechanisms for hallucinations and novel content. The implications extend to enhancing the safety and trustworthiness of AI-generated text, which is increasingly relevant in various domains, including legal and literary contexts.

Authors: Nicholas S. Kersting, Vittorio Castelli, Chieh Ting Yeh, Xinzhu Wang, Saad Taame
Source: arXiv:2605.05103
URL: https://arxiv.org/abs/2605.05103v1