Erik VonKaenel | Searcharxiv

arXiv2026

Semantic Needles in Document Haystacks: Sensitivity Testing of LLM-as-a-Judge Similarity Scoring

We propose a scalable, multifactorial experimental framework that systematically probes LLM sensitivity to subtle semantic changes in pairwise document comparison. We analogize this as a needle-in-a-haystack problem: a single semantically altered sentence (the needle) is embedded within surrounding context (the hay), and we vary the perturbation type (negation, conjunction swap, named entity replacement), context type (original vs. topically unrelated), needle position, and document length across all combinations, testing five LLMs on tens of thousands of document pairs. Our analysis reveals several striking findings. First, LLMs exhibit a within-document positional bias distinct from previously studied candidate-order effects: most models penalize semantic differences more harshly when they occur earlier in a document. Second, when the altered sentence is surrounded by topically unrelated context, it systematically lowers similarity scores and induces bipolarized scores that indicate either very low or very high similarity. This is consistent with an interpretive frame account in which topically-related context may allow models to contextualize and downweight the alterations. Third, each LLM produces a qualitatively distinct scoring distribution, a stable "fingerprint" that is invariant to perturbation type, yet all models share a universal hierarchy in how leniently they treat different perturbation types. Together, these results demonstrate that LLM semantic similarity scores are sensitive to document structure, context coherence, and model identity in ways that go beyond the semantic change itself, and that the proposed framework offers a practical, LLM-agnostic toolkit for auditing and comparing scoring behavior across current and future models.

cs.CL↗

arXiv2021

Selection of inverse gamma and half-t priors for hierarchical models: sensitivity and recommendations

While the importance of prior selection is well understood, establishing guidelines for selecting priors in hierarchical models has remained an active, and sometimes contentious, area of Bayesian methodology research. Choices of hyperparameters for individual families of priors are often discussed in the literature, but rarely are different families of priors compared under similar models and hyperparameters. Using simulated data, we evaluate the performance of inverse gamma and half-$t$ priors for estimating the standard deviation of random effects in three hierarchical models: the 8-schools model, a random intercepts longitudinal model, and a simple multiple outcomes model. We compare the performance of the two prior families using a range of prior hyperparameters, some of which have been suggested in the literature, and others that allow for a direct comparison of pairs of half-$t$ and inverse-gamma priors. Estimation of very small values of the random effect standard deviation led to convergence issues especially for the half-$t$ priors. For most settings, we found that the posterior distribution of the standard deviation had smaller bias under half-$t$ priors than under their inverse-gamma counterparts. Inverse gamma priors generally gave similar coverage but had smaller interval lengths than their half-$t$ prior counterparts. Our results for these two prior families will inform prior specification for hierarchical models, allowing practitioners to better align their priors with their respective models and goals.

stat.ME↗