TASK Quarterly

Efficiency of an unsupervised machine learning approach for superatomic cluster assessment

Sümeyye Atmaca — 2026-05-25

Superatoms are of broad interest in materials science due to their high tunability of electronic properties upon structural modification. The quantum-mechanical (QM) descriptors estimated in our previous work (Sikorska C, Puzyn T, Nanotechnology (2015), 26: 455702) have been used to explore structure-property relationships in superatom-like fullerenes. The structural similarity among fullerene derivatives has been investigated using principal component analysis (PCA) and two-way Hierarchical Cluster Analysis (t-HCA) based on these QM descriptors, demonstrating how electronic structure parameters and geometrical features influence fullerene cluster properties. This unsupervised machine learning approach highlights that descriptors derived from quantum-mechanical calculations enable distinguishing groups of structurally similar compounds for which we can assume similar values of selected physicochemical properties. In addition, the use of computational methods not only reduces the time and costs of research but also the amount of waste generated during experimental analyses. Hence, the research described has significant social and economic significance. At the same time, our results provide a framework for understanding structure-property relationships in nanomaterials that can be used in the future to define new Quantitative Structure-Property Relationship (QSPR) models for predicting physicochemical properties of fullerene-based materials directly from the fullerene structure.

Game of Questions: An automated method for unconventional evaluation of Large Language Models

Przemysław Świat — 2026-05-25

The rapid advancement of Large Language Models (LLMs) has created a need for methods to evaluate their performance, particularly in assessing their domain-specific knowledge and the ability to apply such knowledge in reasoning tasks. Current benchmarks often require substantial manual effort for test case construction and answer scoring. We address this limitation by providing a robust, automatic evaluation method that relies only on unstructured domain text. We introduce the Game of Questions, a method that allows the model's knowledge to be tested via an interaction with another model, inspired by the popular web-based game Akinator. The approach requires minimal input from the evaluator and no prepared questions, making it convenient to apply.

Intrinsic Resonance in Spiking Neural Networks: A Conformable Fractional Mass-Spring-Damper Model for Temporal Feature Extraction

Basem Ajarmah — 2026-05-25

This paper addresses a critical limitation in Spiking Neural Networks (SNNs): the lack of inherent temporal depth in standard integer-order Leaky Integrate-and-Fire (LIF) models. We propose a new neuron model based on the Conformable Fractional Mass-Spring-Damper (MSD) system. Unlike classical models that simply decay over time, our model uses the conformable fractional derivative to give the membrane potential physical momentum and natural resonance. We provide a complete mathematical derivation, clear specification of the learning method, and thorough experimental validation including single-neuron dynamics and full network evaluation on the DVS128 Gesture dataset. Our results show that MSD-based SNNs achieve 93.7% accuracy, significantly outperforming LIF-based baselines (87.3%). The improvement comes from the model's ability to bridge temporal gaps in event-based data through sustained resonant dynamics. The MSD neuron produces 5 autonomous spikes after stimulus offset compared to only 1 for LIF, demonstrating true resonance. Parameter analysis shows stable convergence with mass decreasing by 3%, damping increasing by 20%, and stiffness increasing by 10% during training. Extensive experiments confirm that both fractional orders and resonance contribute to the performance gain. The proposed framework offers a computationally efficient alternative to traditional fractional calculus while providing better temporal feature extraction for neuromorphic computing applications.