Concept Fields: Measuring Hallucination and Novelty in Text Corpora
A new method called Concept Fields measures text coherence and novelty by analyzing sentence transitions in embedding space. Introduced in arXiv paper 2605.05103, it computes a local drift field with pointwise uncertainty from deltas between consecutive sentences. A score ζ quantifies agreement with the field using mean absolute z-distance. The approach is black-box and corpus-attributable, supported by a Vector Sequence Database (VSDB). Evaluated on hallucination detection in the U.S. Code of Federal Regulations and novelty detection in Project Gutenberg, it uses controlled LLM-generated rewrites.
Key facts
- Concept Fields are local drift fields with pointwise uncertainty estimated from sentence-embedding deltas.
- Score ζ measures mean absolute z-distance between observed delta and local Gaussian estimate.
- The method is black-box and corpus-attributable.
- A Vector Sequence Database (VSDB) stores embeddings with sequence-position and next-delta metadata.
- Evaluated on U.S. Code of Federal Regulations for hallucination detection.
- Evaluated on Project Gutenberg for novelty detection.
- Uses controlled LLM-generated rewrites.
- Paper is arXiv:2605.05103.
Entities
Institutions
- arXiv
- U.S. Code of Federal Regulations
- Project Gutenberg