analiza aiqbench
The paper proposes "Concept Fields," a lightweight method that scores whether a sentence transition is grounded in a text corpus, using sentence-embedding drift statistics and a new Vector Sequence Database. It reports selective classification results on U.S. federal regulations and Project Gutenberg. For readers tracking LLM reliability, the main question is whether this fast, model-agnostic, interpretable signal holds up against LLM-as-judge and white-box detectors on real workloads. The evidence here is limited to two corpora, so its value as a complement to existing methods is still unproven.
Text Corpora as Concept Fields: Providing Explainable Groundedness
arXiv:2605.05103v4 Announce Type: replace-cross Abstract: We introduce the \textbf{Concept Field} of a text corpus: a local drift field with pointwise uncertainty, estimated in sentence-embedding space from the deltas between consecutive sentences. Given a candidate sentence transition, we score its agreement with the field by $\zeta$, the mean absolute z-distance between the observed delta and the field's local statistical estimate, giving a quantitative measure of groundedness. The score is model agnostic and corpus-attributable, providing traces to corpus semantics either supporting or refuting groundedness. We drive the computation with the introduction of a \textbf{Vector Sequence Database (VSDB)} that stores embeddings together with sequence-position and next-delta metadata. We demonstrate this approach on two large-scale settings: groundedness detection over the U.S. Code of Federal Regulations, and ungroundedness (novelty) detection over Project Gutenberg, where we show Concept Fields achieve strong selective classification performance under a grounded / ungrounded / unsure triage policy. Unlike retrieval-centric baselines, the resulting coverage-risk behavior shows similar empirical coverage-performance trends in the two evaluated domains. Concept Fields provide a fast, lightweight, and interpretable signal for groundedness, complementary to LLM-as-judge and white-box detectors.
Abstract:We introduce the \textbf{Concept Field} of a text corpus: a local drift field with pointwise uncertainty, estimated in sentence-embedding space from the deltas between consecutive sentences. Given a candidate sentence transition, we score its agreement with the field by $\zeta$, the mean absolute z-distance between the observed delta and the field's local statistical estimate, giving a quantitative measure of groundedness. The score is model agnostic and corpus-attributable, providing traces to corpus semantics either supporting or refuting groundedness. We drive the computation with the introduction of a \textbf{Vector Sequence Database (VSDB)} that stores embeddings together with sequence-position and next-delta metadata. We demonstrate this approach on two large-scale settings: groundedness detection over the U.S. Code of Federal Regulations, and ungroundedness (novelty) detection over Project Gutenberg, where we show Concept Fields achieve strong selective classification performance under a grounded / ungrounded / unsure triage policy. Unlike retrieval-centric baselines, the resulting coverage-risk behavior shows similar empirical coverage-performance trends in the two evaluated domains. Concept Fields provide a fast, lightweight, and interpretable signal for groundedness, complementary to LLM-as-judge and white-box detectors.
Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Computers and Society (cs.CY)
arXiv:2605.05103 [cs.CL]
(or arXiv:2605.05103v4 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2605.05103 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 276 cuvinte
Subiecte conexe
A Theory of Post-hoc Debate Judgement
arXiv · 21 aug. 2026
Google Research Introduces Retrieve-for-Train (R4T): An RL-Compiled Diffusion Retriever for 12× to 20× Faster Query Fan-Out
MarkTechPost · 17 sept. 2026
Single-Round Vector RAG vs an LLM-Compiled Wiki: A Preregistered Comparison on a Small Multi-Domain Research Corpus
arXiv · 18 aug. 2026
Adapting Vision-Language Models for Human-Readable XAI in Industrial Object Detection
arXiv · 29 sept. 2026