Selected work

Representative Papers

Comment on: Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Tasks

Dec 29, 2025

This study critically evaluates the methodological rigor and interpretive validity of Kosmyna et al.’s research on cognitive debt induced by ChatGPT use. Through systematic literature analysis, automated NLP pipelines, and reproducibility assessment of EEG data, the work identifies significant shortcomings in the original study, including insufficient sample size, opaque EEG preprocessing procedures, non-reproducible analytical workflows, and inconsistent reporting of results. This project represents the first comprehensive examination of methodological flaws and interpretive biases in research on the cognitive impacts of AI-assisted writing. It calls for a more cautious interpretation of claims regarding “cognitive debt” and provides concrete recommendations for establishing stricter scientific standards in this emerging field.

1 citations1 influentialRead paper

Algebraic and Arithmetic Attributes of Hypergeometric Functions in SageMath

Feb 04, 2026

This study investigates the algebraicity and arithmetic properties of hypergeometric functions over the rational numbers, finite fields, and p-adic fields. Leveraging the SageMath computer algebra system, the work integrates techniques from algebraic number theory, finite field theory, and p-adic analysis to systematically implement, for the first time in an open-source framework, algorithms capable of determining algebraicity, computing valuations, and solving for minimal polynomials in positive characteristic. This implementation fills a critical gap in existing computational toolchains by enabling uniform arithmetic analysis of hypergeometric functions across multiple number-theoretic domains, thereby substantially enhancing SageMath’s capacity for algebraic manipulation of such functions.

1 citationsRead paper

Algorithms for Algebraic and Arithmetic Attributes of Hypergeometric Functions

Jan 22, 2026

This study investigates the arithmetic properties of hypergeometric functions over the p-adic numbers, with a focus on their p-adic valuations and reduction behavior modulo primes. Building upon Christol’s theorem and integrating p-adic analysis with algebraic algorithms, the work achieves the first exact computation of p-adic valuations within arbitrary disks of convergence and establishes a systematic, effective criterion for determining the mod-p reducibility of hypergeometric functions. Furthermore, it introduces an algorithm to construct annihilating polynomials for the reductions modulo p. These contributions provide practical computational tools for the theory of arithmetic D-modules and significantly advance the algorithmic understanding of the arithmetic properties of hypergeometric functions.

1 citationsRead paper

Do LLM Self-Explanations Help Users Predict Model Behavior? Evaluating Counterfactual Simulatability with Pragmatic Perturbations

Jan 07, 2026arXiv.org

This study investigates whether self-explanations generated by large language models (LLMs) can improve the accuracy of human and LLM predictions regarding the models’ behavior in counterfactual scenarios. To this end, we introduce a novel integration of pragmatic perturbations with a counterfactual simulatability framework to construct test cases, and conduct a systematic evaluation using chain-of-thought and post-hoc explanation generation, joint human–LLM assessments, and qualitative analysis of free-text responses. Our findings demonstrate that self-explanations significantly enhance prediction accuracy, though this effect is moderated by the choice of perturbation strategy and the evaluators’ reasoning capabilities. Further analysis of user-generated rationales corroborates the constructive role of explanations in shaping human judgment.

1 citationsRead paper

Theory-to-Practice Gap for Neural Networks and Neural Operators

Mar 23, 2025

This work investigates the sampling complexity of ReLU neural networks and neural operators when learning mappings belonging to specific approximation spaces, focusing on the “theory–practice gap” between theoretically optimal convergence rates and empirically attainable ones. Within a unified $L^p$ framework, we systematically extend this gap to infinite-dimensional operator learning—its first such treatment—using Bochner integration and approximation-theoretic tools to rigorously establish that the optimal sampling convergence rate is fundamentally limited by the $1/p$-order Monte Carlo rate. Our analysis encompasses mainstream architectures including DeepONet and the Fourier Neural Operator (FNO), and improves upon existing convergence upper bounds. Key contributions are: (1) a fine-grained separation of parameter complexity from sampling complexity; (2) identification of an intrinsic Monte Carlo–type convergence bottleneck in infinite-dimensional operator learning; and (3) provision of tight theoretical limits to guide principled neural operator design.

1 citationsRead paper

Recent publications

Latest Papers

LMM Modality Transfer: A Pre-requisite for Autonomous GIS Agents

Aug 07, 2026

This study addresses the challenge of ineffective alignment between image and text modalities in large models when applied to geographic information system (GIS) tasks, which hinders progress in automated spatial analysis. To evaluate cross-modal spatial information transfer, the authors propose an end-to-end modality translation framework: a large multimodal model (LMM) first generates textual descriptions from colored grid images, and a second LMM reconstructs the original spatial image from this description. This approach introduces spatial information theory into LMM evaluation for the first time, demonstrating that multimodal consistency is essential for robust geospatial understanding. Experimental results reveal that even state-of-the-art LMMs perform poorly on simple grid-based scenes, highlighting significant deficiencies in current models’ capacity for spatial semantic alignment.

0 citationsRead paper

A Symbolic Execution Framework for Symbolic Timing Analysis of Digital Integrated Circuits

Aug 03, 2026

Traditional simulation-based dynamic timing analysis struggles to balance accuracy and efficiency, and existing gate delay models lack sufficient expressiveness to enable precise, exhaustive path delay analysis for digital circuits. This work proposes a symbolic execution framework integrated with an analytical gate delay model that automatically generates symbolic delay expressions for all paths under a given input transition ordering. For the first time, it incorporates an analytical delay model accounting for both drafting effects and multi-input switching into symbolic execution. By employing a path-sensitive, goal-directed inference mechanism together with symbolic pruning strategies, the approach significantly enhances the completeness and precision of timing analysis while effectively mitigating the combinatorial explosion problem.

0 citationsRead paper

Blind Source Separation Can Distort Behavior and Connectivity Analyses of Calcium Transients

Aug 01, 2026

Blind source separation (BSS) is commonly employed as a neutral denoising step in calcium transient analysis, yet it may inadvertently interfere with behavioral decoding and causal connectivity inference. This study systematically evaluates four BSS algorithms—FastICA, Infomax, SOBI, and JADE—on both synthetic data and real zebrafish v2a-RSN recordings, benchmarking them against PCA and low-pass filtering using conditional Granger causality (c-GC/c-GC*) for connectivity inference. Results indicate that while BSS occasionally enhances behavioral decoding performance, such improvements are inconsistent. More critically, BSS consistently induces significant densification of inferred causal graphs, reducing graph recovery F1 scores to zero in synthetic datasets. These findings demonstrate that BSS acts as an interventional rather than an unbiased preprocessing step, thereby challenging prevailing analytical paradigms in the field.

0 citationsRead paper

Three-Photon Bayesian Imaging of Ortho-Positronium

Jul 30, 2026

This study addresses the underutilization of three-photon decay signals from positronium—particularly ortho-positronium—in conventional PET imaging, which discards potentially valuable microenvironmental biomarkers by considering only two-photon coincidence events. To harness this untapped information, the authors propose TRIO, a novel algorithm that, for the first time, integrates quantum electrodynamics (QED)-derived physical priors into a Bayesian maximum a posteriori framework. TRIO jointly exploits time-of-flight, energy constraints, and trilateration to achieve event-level three-photon reconstruction. Notably, the method is compatible with standard radiotracers such as ¹⁸F, thereby eliminating the dependency on specialized isotopes traditionally required for positronium lifetime imaging. In simulations based on the Siemens Biograph Quadra, TRIO reduces the average localization error to 1.62 cm—approximately twofold better than time-only methods and nearly an order of magnitude superior to energy-only approaches—while remaining adaptable to existing TOF-PET systems.

0 citationsRead paper

Send and Pretend: Exploiting Transcript Consistency Issues in End-to-End Encrypted Group Chats

Jul 29, 2026

This study addresses a critical yet overlooked vulnerability in mainstream end-to-end encrypted (E2EE) group messaging systems: the lack of transcript consistency guarantees. Malicious participants can exploit protocol fallbacks, pairwise channels, and other mechanisms to deliver inconsistent messages to different recipients, thereby undermining group consensus without detection. The work presents the first systematic analysis of this issue, introducing a cross-platform attack model that uncovers novel privacy risks such as device fingerprint leakage. Through protocol reverse engineering, message path tracing, and UI behavior testing across real-world E2EE applications, the authors identify multiple inconsistency vectors and demonstrate practical attacks—including vote manipulation, moderation evasion, and social engineering. Finally, they propose lightweight mitigation mechanisms and user interface warning strategies that significantly enhance transcript consistency in group chats.

0 citationsRead paper