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

Source: arXiv - 2605.05103v1

Overview

The paper proposes a new way to treat any text collection as a Concept Field—a statistical map that captures how ideas naturally flow from one sentence to the next. By measuring how well a new sentence transition aligns with this field, the authors can flag whether a statement is likely “hallucinated” (i.e., unsupported by the source material) or genuinely novel. The approach works without peeking inside the language model, making it attractive for black‑box safety checks and novelty detection in real‑world pipelines.

Key Contributions

• Concept Field definition – a local drift field in sentence‑embedding space, modeled as a Gaussian distribution of sentence‑to‑sentence deltas.
• ζ‑score – a simple, interpretable metric (mean absolute z‑distance) that quantifies agreement between a candidate transition and the field.
• Vector Sequence Database (VSDB) – an efficient index that stores embeddings together with positional and delta metadata, enabling fast field look‑ups.
• Domain‑agnostic evaluation – applied to (1) hallucination detection in regulatory text (U.S. Code of Federal Regulations) and (2) novelty detection in literary works (Project Gutenberg).
• Probabilistic triage policy – a three‑way decision (grounded / ungrounded / unsure) that yields comparable coverage‑risk curves across domains, unlike retrieval‑centric baselines.
• Qualitative analysis tools – divergence and curl of the field surface semantic patterns (e.g., logical “sources” and “sinks”) that can guide hypothesis generation.

Methodology

1. Sentence Embedding – Each sentence is encoded with a dense vector (e.g., using a pre‑trained transformer encoder).

2. Delta Computation – For every adjacent pair in the corpus, the vector difference (Δ) is recorded, producing a cloud of deltas that represent the natural “direction” of discourse.

3. Local Gaussian Estimation – Around any query point (the embedding of a candidate sentence), the surrounding deltas are modeled as a multivariate Gaussian (mean μ, covariance Σ).

4. ζ‑Score Calculation – The observed delta of the candidate transition is compared to the local Gaussian via a mean absolute z‑distance:

   $$
   ζ = \frac{1}{d}\sum_{i=1}^{d}\big| \frac{Δ_i - μ_i}{\sqrt{Σ_{ii}}} \big|
   $$

   Lower ζ indicates high conformity to the corpus; higher ζ suggests a potential hallucination or novelty.

5. VSDB Indexing – Embeddings, their positions, and the next‑sentence deltas are stored in a vector‑sequence database, allowing sub‑linear retrieval of the local neighborhood needed for the Gaussian estimate.

6. Triaging – Thresholds on ζ define three zones: “grounded” (ζ below low threshold), “un‑grounded” (ζ above high threshold), and “unsure” (in‑between).

The whole pipeline runs outside any LLM, requiring only embeddings and the VSDB, which makes it lightweight and easy to plug into existing systems.

Results & Findings

Key takeaways

• The ζ‑score provides a probabilistic interpretation (e.g., “the observed transition is 3σ away from the corpus norm”).
• Performance holds up even when the underlying LLM is a black box, showing that the field captures corpus‑level semantics robustly.
• Divergence/curl visualizations reveal interpretable structures such as “topic attractors” (high‑divergence points) and “topic sinks” (low‑curl regions), useful for exploratory analysis.

Practical Implications

• Safety‑first LLM pipelines – Insert a Concept Field check before exposing model outputs to end‑users; the triage can automatically block high‑ζ responses or flag them for human review.
• Content moderation & plagiarism detection – Use ζ to spot novel or out‑of‑domain text in user‑generated content without needing to store the entire source corpus.
• Rapid domain adaptation – Build a VSDB on a small, domain‑specific corpus (e.g., internal API docs) and instantly obtain a groundedness filter for any downstream LLM.
• Lightweight alternative to retrieval‑augmented generation – Instead of pulling full passages, the field supplies a scalar confidence signal that can be combined with retrieval scores for richer ranking.
• Explainability – Because each ζ score traces back to nearby corpus sentences, developers can surface the “supporting” sentences to justify why a claim is considered grounded or not.

Limitations & Future Work

• Embedding dependence – The quality of the field hinges on the chosen sentence encoder; biased or low‑capacity embeddings could mischaracterize drift.
• Local Gaussian assumption – Real discourse may exhibit multimodal or heavy‑tailed delta distributions, which a single Gaussian cannot capture.
• Scalability of VSDB – While sub‑linear, building and updating the VSDB for continuously growing corpora remains an engineering challenge.
• Domain transfer – The paper shows promising cross‑domain consistency, but systematic evaluation on highly technical domains (e.g., medical or code) is still pending.
• Future directions suggested include richer density models (mixture of Gaussians or normalizing flows), incremental VSDB updates, and integrating ζ as a soft regularizer during LLM fine‑tuning.

Bottom line: Concept Fields turn any text collection into a statistical “map of meaning flow,” offering developers a fast, model‑agnostic signal to detect hallucinations and novelty. With modest engineering overhead (embedding extraction + VSDB), this technique can be layered onto existing LLM services to boost safety, transparency, and domain awareness.

Authors

• Nicholas S. Kersting
• Vittorio Castelli
• Chieh Ting Yeh
• Xinzhu Wang
• Saad Taame

Paper Information

• arXiv ID: 2605.05103v1
• Categories: cs.CL, cs.AI, cs.CY
• Published: May 6, 2026
• PDF: Download PDF