-
Beyond Document Grounding: Span-Level Hallucination Detection over Code, Tool Output, and Documents
Authors:
Ádám Kovács,
Bowei He,
Xue Liu,
István Boros,
Szilveszter Tóth,
Gábor Recski
Abstract:
Hallucination detection for retrieval-augmented generation (RAG) is usually evaluated on natural-language document evidence. However, grounded generation systems increasingly rely on structured inputs: source code, developer-tool output, markdown documents, tables, and repository metadata. We introduce a unified benchmark for span-level hallucination detection over code, tool output, structured do…
▽ More
Hallucination detection for retrieval-augmented generation (RAG) is usually evaluated on natural-language document evidence. However, grounded generation systems increasingly rely on structured inputs: source code, developer-tool output, markdown documents, tables, and repository metadata. We introduce a unified benchmark for span-level hallucination detection over code, tool output, structured documents, and existing natural-language RAG datasets. The benchmark is built by starting from grounded correct answers, injecting localized hallucinations with exact character labels, and validating the code test split with evidence-based review. Our fine-tuned Qwen3.5-2B detector reaches 0.689 span-F1 on the unified test set and 0.60 on the code-agent source, where it substantially outperforms LettuceDetect-large (0.17) and the strongest zero-shot LLM judges we evaluated (at most 0.22). The same model remains competitive on established natural-language benchmarks, with 81.8 RAGTruth example-F1 and 0.724 English PsiloQA IoU.
△ Less
Submitted 1 July, 2026;
originally announced July 2026.
-
Feasibility demonstration of continuous signal-based neutron noise measurements by experiments and simulations
Authors:
Máté István Boros,
Máté Szieberth,
Gergely Klujber,
Imre Pázsit,
István Barth,
Yasunori Kitamura,
Tsuyoshi Misawa
Abstract:
Neutron noise methods are used to determine kinetic parameters such as the prompt neutron decay constant, but traditional pulse-counting suffers from dead-time and pile-up at high detection rates. Recent theory shows that analysing the continuous detector current can avoid these limitations if pulse-shape effects are properly treated. This work presents a feasibility study of continuous-signal neu…
▽ More
Neutron noise methods are used to determine kinetic parameters such as the prompt neutron decay constant, but traditional pulse-counting suffers from dead-time and pile-up at high detection rates. Recent theory shows that analysing the continuous detector current can avoid these limitations if pulse-shape effects are properly treated. This work presents a feasibility study of continuous-signal neutron noise analysis based on simulations and experiments performed at two research reactors. The stochastic model of the detector current is applied to derive Rossi- and Feynman-type formulations, and pulse-shape distortions are mitigated using detector pairs or by deconvolving the average pulse-shape through inverse Fourier and Wiener filtering. Simulations demonstrate accurate $α$-parameter estimation at count rates where pulse-counting becomes unusable, and enable evaluation of significantly higher $α$ values. Measurements at KUCA and BME TR confirm that continuous and deconvolved signals provide unbiased results despite dead-time and electronic artifacts, establishing the method as a practical alternative for high-rate reactor noise diagnostics.
△ Less
Submitted 9 June, 2026;
originally announced June 2026.
-
ACL-Verbatim: hallucination-free question answering for research
Authors:
Gábor Recski,
Szilveszter Tóth,
Nadia Verdha,
István Boros,
Ádám Kovács
Abstract:
Academic researchers need efficient and reliable methods for collecting high-quality information from trusted sources, but modern tools for AI-assisted research still suffer from the tendency of Large Language Models (LLMs) to produce factually inaccurate or nonsensical output, commonly referred to as hallucinations. We apply the extractive question answering system VerbatimRAG to research papers…
▽ More
Academic researchers need efficient and reliable methods for collecting high-quality information from trusted sources, but modern tools for AI-assisted research still suffer from the tendency of Large Language Models (LLMs) to produce factually inaccurate or nonsensical output, commonly referred to as hallucinations. We apply the extractive question answering system VerbatimRAG to research papers in the ACL Anthology, directly mapping user queries to verbatim text spans in retrieved documents. We contribute a novel ground truth dataset for the task of mapping user queries to relevant text spans in research papers, and use it to train and evaluate a variety of extractive models. Human annotation is performed by NLP researchers and is based on synthetic user queries generated using a custom pipeline based on the ScIRGen methodology, paired with chunks of research papers retrieved by VerbatimRAG. On this benchmark, a 150M-parameter ModernBERT token classifier trained on silver supervision from our pipeline achieves the best word-level F1 (53.6), ahead of the strongest evaluated LLM extractor (48.7).
△ Less
Submitted 20 May, 2026;
originally announced May 2026.
-
Benchmark for numerical solutions of flow in heterogeneous groundwater formations
Authors:
Cristian D. Alecsa,
Imre Boros,
Florian Frank,
Peter Knabner,
Mihai Nechita,
Alexander Prechtel,
Andreas Rupp,
Nicolae Suciu
Abstract:
This article presents numerical investigations on accuracy and convergence properties of several numerical approaches for simulating steady state flows in heterogeneous aquifers. Finite difference, finite element, discontinuous Galerkin, spectral, and random walk methods are tested on one- and two-dimensional benchmark flow problems. Realizations of log-normal hydraulic conductivity fields are gen…
▽ More
This article presents numerical investigations on accuracy and convergence properties of several numerical approaches for simulating steady state flows in heterogeneous aquifers. Finite difference, finite element, discontinuous Galerkin, spectral, and random walk methods are tested on one- and two-dimensional benchmark flow problems. Realizations of log-normal hydraulic conductivity fields are generated by Kraichnan algorithms in closed form as finite sums of random periodic modes, which allow direct code verification by comparisons with manufactured reference solutions. The quality of the methods is assessed for increasing number of random modes and for increasing variance of the log-hydraulic conductivity fields with Gaussian and exponential correlation. Experimental orders of convergence are calculated from successive refinements of the grid. The numerical methods are further validated by comparisons between statistical inferences obtained from Monte Carlo ensembles of numerical solutions and theoretical first-order perturbation results. It is found that while for Gaussian correlation of the log-conductivity field all the methods perform well, in the exponential case their accuracy deteriorates and, for large variance and number of modes, the benchmark problems are practically not tractable with reasonably large computing resources, for all the methods considered in this study.
△ Less
Submitted 3 March, 2020; v1 submitted 25 November, 2019;
originally announced November 2019.
-
New optimization algorithms for neural network training using operator splitting techniques
Authors:
Cristian Daniel Alecsa,
Titus Pinta,
Imre Boros
Abstract:
In the following paper we present a new type of optimization algorithms adapted for neural network training. These algorithms are based upon sequential operator splitting technique for some associated dynamical systems. Furthermore, we investigate through numerical simulations the empirical rate of convergence of these iterative schemes toward a local minimum of the loss function, with some suitab…
▽ More
In the following paper we present a new type of optimization algorithms adapted for neural network training. These algorithms are based upon sequential operator splitting technique for some associated dynamical systems. Furthermore, we investigate through numerical simulations the empirical rate of convergence of these iterative schemes toward a local minimum of the loss function, with some suitable choices of the underlying hyper-parameters. We validate the convergence of these optimizers using the results of the accuracy and of the loss function on the MNIST, MNIST-Fashion and CIFAR 10 classification datasets.
△ Less
Submitted 21 March, 2020; v1 submitted 29 April, 2019;
originally announced April 2019.