Cryptoasset systems often bind cryptographic key control to financial control: losing a wallet seed, custody share, hardware device, or smart-account credential can remove spend authority, while compromised recovery can enable theft. Existing work treats recovery through separate vocabularies–key backup, secret sharing, account recovery, credential re-issuance, social recovery, and asset migration–making mechanisms and tradeoffs difficult to compare. This paper presents a Systematization of Knowledge (SoK) on cryptographic key recovery for cryptoasset custody and financial technologies. Starting from a 118-paper systematic-review discovery corpus, we derive a 77-paper synthesis corpus and code each retained system in a master matrix covering recovered objects, recovery semantics, mechanisms, enrollment and storage, authorization, trust placement, failure events, post-recovery state, validation evidence, deployment status, privacy, usability, and limitations. The matrix supports an axis-first taxonomy that separates secret-restoring, hybrid, control-restoring, forensic/extractive, and framework-oriented recovery. Our central observation is that recovery is not a single operation: systems may reconstruct an original secret, regenerate a seed, restore a share, reissue a credential, migrate signing authority, restore account control, move assets, or extract forensic artifacts. We derive a generalized construction model, check it against production-facing designs, and identify six findings: recovery semantics are heterogeneous; recovery shifts trust; liveness improvements create abuse paths; post-recovery lifecycle management is uneven; protocol evidence outpaces user evidence; and recovery metadata remains underprotected. These gaps motivate a research agenda for recovery-aware financial technologies.

Hyperdimensional Computing (HDC) is a robust computational framework inspired by human cognition characterized by simple and efficient operations within high-dimensional vector spaces. Quantum-enhanced Hyperdimensional Computing (QeHDC) extends classical HDC by leveraging quantum mechanical properties to enhance computational efficiency. In this paper, we propose a novel Quantum HDC framework featuring a one-pass training method, leveraging sinusoidal and quantum encoding to project classical data into quantum amplitude states efficiently. Our framework introduces an innovative reference-state-based quantum binding operation realized via quantum circuits. Furthermore, we propose a density-matrix-based superclass generation strategy employing eigenvalue decomposition to extract critical quantum state features effectively, enabling a more accurate and robust class representation. Experimental evaluations conducted on standard benchmark datasets demonstrate our approach's superior performance, robustness to noise, and computational feasibility compared to traditional classical and existing quantum-enhanced approaches. The results highlight the practical benefits and potential of Quantum HDC for quantum-enhanced classification tasks and pave the way for future advancements in quantum-inspired computational paradigms.

This paper presents an explainable triple-agent AI workflow for OECD-oriented tax risk assessment using public Country-by-Country Reporting (CbCR) data. The study addresses the lack of scalable and reproducible analytical approaches for interpreting multinational profit allocation. The novelty lies in integrating LLM-based risk evaluation, automated Requests for Explanation, and taxpayer response templates into a traceable compliance workflow tested on the AITAX13 database.

Agent skills are widely supported by major agentic frameworks and perform well with proprietary models, yet their effectiveness for small and medium-sized open source language models (270 M-80B) remains underexplored. We systematically study the Skill paradigm in resource-constrained industrial settings, where reliance on proprietary APIs is impractical due to data security and budget constraints. Across two open-source tasks and a real-world insurance claims classification task, we find that very small models struggle with reliable skill selection, while models around 30B-80B benefit substantially. Thinking variants do not show major levels of improvement from skills, also considering GPU usage increases due to overthinking. These findings reveal a trade-off between GPU cost and agent performance, and provide actionable insights for effective Skill configuration and SLM deployment in real world settings.

We present a geometric framework for analysing multi-head attention in large language models (LLMs). Without altering the mechanism, we view standard attention through a top-N selection lens and study its behaviour directly in value-state space. We define geometric metrics - Precision, Recall, and F-score - to quantify separability between selected and non-selected tokens, and derive non-asymptotic bounds with explicit dependence on dimension and margin under empirically motivated assumptions (stable value norms with a compressed sink token, exponential similarity decay, and piecewise attention weight profiles). The theory predicts a small-N operating regime of strongest non-trivial separability and clarifies how sequence length and sink similarity shape the metrics. Empirically, across LLaMA-2-7B, Gemma-7B, and Mistral-7B, measurements closely track the theoretical envelopes: top-N selection sharpens separability, sink similarity correlates with Recall. We also found that in LLaMA-2-7B heads specialize into three regimes - Retriever, Mixer, Reset - with distinct geometric signatures. Overall, attention behaves as a structured geometric classifier with measurable criteria for token selection, offering head level interpretability and informing geometry-aware sparsification and design of attention in LLMs.

Recent research on time-series foundation models (TSFMs) has underscored the scarcity of real-world data, often supplemented with synthetic sources in existing datasets, whose generalizability remains however debated. As such, in this work, we propose a novel benchmarking approach: in particular, we aim at building a curated dataset reflecting real world physical temporal dynamics, extracting temporal signals from real-world videos using optical flow. As such, we introduce REAL-V-TSFM, a novel dataset designed to capture rich and diverse time series derived from real-world videos. Experimental results on state-of-the-art TSFMs under zero-shot forecasting show that, despite strong performance on conventional benchmarks, these models exhibit performance degradation on the proposed dataset, suggesting limited generalizability to novel datasets. These findings underscore the need for novel approaches to acquiring time series data and highlight the lack of universality in recent TSFMs, while further validating the effectiveness of our video-based time series data extraction pipeline.

Hyperdimensional Computing (HDC) is a brain-inspired computational paradigm that utilizes high-dimensional vector spaces to represent, bind, and process information efficiently. Its robustness, scalability, and adaptability make HDC particularly effective for machine learning tasks such as classification and pattern recognition. In this work, we introduce Quantum-StateHD (QSHD), a novel approach that leverages random quantum states as encoding mechanisms for high-dimensional spaces within the HDC framework. Quantum states, with their inherent properties of high-dimensionality, near-orthogonality, and probabilistic distributions, provide a natural substrate for constructing robust representations. Quantum-StateHD encodes classical data into the coefficients of random quantum states and seamlessly integrates these representations into the HDC pipeline. Experimental results on several benchmark classification datasets demonstrate that Quantum-StateHD achieves competitive or superior performance compared to traditional state-of-the-art HDC methods, showcasing its potential for robust and efficient high-dimensional information processing. This study bridges the gap between quantum information science and HDC, offering a promising direction for hybrid quantum-classical machine learning architectures.

With the emergence of ChatGPT, Transformer models have significantly advanced text classification and related tasks. Decoder-only models such as Llama exhibit strong performance and flexibility, yet they suffer from inefficiency on inference due to token-by-token generation, and their effectiveness in text classification tasks heavily depends on prompt quality. Moreover, their substantial GPU resource requirements often limit widespread adoption. Thus, the question of whether smaller language models are capable of effectively handling text classification tasks emerges as a topic of significant interest. However, the selection of appropriate models and methodologies remains largely underexplored. In this paper, we conduct a comprehensive evaluation of prompt engineering and supervised fine-tuning methods for transformer-based text classification. Specifically, we focus on practical industrial scenarios, including email classification, legal document categorization, and the classification of extremely long academic texts. We examine the strengths and limitations of smaller models, with particular attention to both their performance and their efficiency in Video Random-Access Memory (VRAM) utilization, thereby providing valuable insights for the local deployment and application of compact models in industrial settings.

Several distributed systems based on unstructured p2p networks, such as blockchains, rely on underlying protocols to disseminate messages in a fast and reliable way. As the state-of-the-art for message dissemination in blockchains, GossipSub guarantees delivery and resilience against attacks and byzantine faults by scaling pubsub dissemination without exceeding bandwidth or overloading peers. Although GossipSub relies heavily on the way its mesh is constructed, there is little insight into how different configuration parameters impact the overall performance of the system. This study analyzes the relationships between the configuration and the performance of GossipSub from a causal point of view using the concrete case of the XRPL. By employing graphical causal methods to investigate the strength of those connections, this study goes towards the direction of finding the best configuration for GossipSub for different domains, without the need for excessive empirical tests.

The rise of decentralized finance has brought a vast range of opportunities to the blockchain space and many risks. This paper tackles the challenge of detecting malicious smart contracts on Ethereum designed to exploit vulnerabilities and cause financial losses. We present a novel approach for preemptively identifying malicious smart contracts during their deployment stage. For this purpose, we gathered a dataset comprising 161 malicious smart contracts and 5500 benign smart contracts. By introducing and extracting various features related to the deployer, transaction characteristics, and deployment bytecode and selecting the most impactful features, we developed multiple models using different machine learning (ML) classification algorithms, compared them using the set of most impactful features, and selected the most accurate one as our detection model. We compared the model’s performance with a publicly available ML malicious smart contract detection tool to benchmark it. The results demonstrate that our model achieves a superior True Positive Rate while having a lower False Positive Rate. Our model achieved a 79.17

The integration of large language models (LLMs) and vision language models (VLMs) into forensic applications offers unprecedented opportunities for reconstructing crime scenes from textual testimonies. In this work, PTVR-Pipeline is introduced as the first prototype pipeline that transforms textual witness statements into interactive virtual reality (VR) environments. This approach leverages LLMs for semantic extraction of entities and events, combined with generative 3D modeling tools to produce immersive crime scenes in Unity. The pipeline incorporates prompt engineering strategies, JSON-based object encoding, and multi-testimony alignment to ensure coherence and temporal consistency. The experimental evaluation included benchmarking different LLMs for reliability and qualitatively assessing the generated environments. The results demonstrate the feasibility of using LLM-driven extraction and hybrid 3D generation methods to reconstruct crime scenes with sufficient accuracy for immersive exploration. The PTVR-Pipeline constitutes a significant step towards AI-assisted forensic visualization, while acknowledging the limitations of model hallucination and the dependency on the prompt’s quality.

The Federated Byte-level Byte-Pair Encoding (BPE) Tokenizer (FedByteBPE) leverages a Federated Learning (FL) approach for a privacy-preserving approach to train language models tokenizer across distributed datasets. This approach enables entities to train and refine their tokenizer models locally, with vocabulary aggregation performed on a centralized server. This method ensures the creation of a robust, domain-specific tokenizer while preserving privacy. Supported by theoretical analysis and empirical results from experiments on a real-world distributed financial dataset, our findings demonstrate that the federated tokenizer significantly outperforms off-the-shelf and individual local tokenizers in vocabulary coverage. This highlights the potential of federated learning to address training language model tokenizers in a privacy-preserving setting.

Blockchain has revolutionized finance through decentralization, eliminating the need for traditional intermediaries. However, security concerns remain a major barrier to adoption, as DeFi platforms increasingly face targeted attacks. In this paper, we present the first methodology for automatically assessing and quantifying the risk of fund loss in DeFi projects due to smart contract exploits. By analyzing on-chain behaviors that signal potential malicious interactions, our approach assigns a dynamic risk score to DeFi projects over time. Relying solely on on-chain data ensures resistance to data manipulation and enhances the integrity of the assessment. We evaluated 220 compromised and 200 unaffected DeFi projects on multiple EVM-compatible blockchains - including Ethereum, BSC, Polygon, Arbitrum, Optimism, and Fantom - and conducted a comparative risk assessment on these projects. Our findings reveal statistically significant differences in risk scores before attacks compared to a control group without attacks. We anticipated potential threats to 86% of the projects that were later attacked, one day before the incidents, with a precision of 78%.

The Web of Agents (WoA) transforms the document-centric Web into an environment of autonomous agents acting on users' behalf, a vision newly tractable as large language models (LLMs) mature. We argue that across three decades the WoA has undergone a semantic-effort migration in chronological order: from platform-side coordination (Multi-Agent Systems, Generation I), through data-side annotation (Semantic Web, Generation II), to model-side interpretation (LLM-era, Generation III). The central Gen II → Gen III transition within this trajectory, which we call the semantics-in-data → semantics-in-models shift, is predictive: each generation's failure modes and current open problems follow from where that generation located its semantic effort. The survey makes five contributions: (i) a unified evolutionary narrative spanning 1990–2026; (ii) a four-dimensional comparative framework (semantic foundation, communication paradigm, locus of intelligence, discovery mechanism) applied uniformly across all three generations; (iii) classification of sixteen representative systems on these dimensions, including hybrid LLM–knowledge-graph and computer-use agents; (iv) coverage of the November 2024–August 2026 institutional convergence (Linux Foundation's Agentic AI Foundation, A2A v1.0, MCP November 2024 launch and November 2025 specification, Visa/Mastercard/Stripe payment-network protocols, EU AI Act phased enforcement, the NIST AI Agent Standards Initiative, International AI Safety Report 2026); and (v) seven named lessons grounded in cross-generational evidence paired with seven generation-invariant challenges that persist regardless of which protocol prevails. Further progress depends less on protocol design than on the socio-technical infrastructure now being assembled by standards bodies, regulators, and commercial payment networks.

Cloud cover in multispectral imagery (MSI) poses significant challenges for early season crop mapping, as it leads to missing or corrupted spectral information. Synthetic aperture radar (SAR) data, which is not affected by cloud interference, offers a complementary solution, but lack sufficient spectral detail for precise crop mapping. To address this, we propose a novel framework, Time-series MSI Image Reconstruction using Vision Transformer (ViT), to reconstruct MSI data in cloud-covered regions by leveraging the temporal coherence of MSI and the complementary information from SAR from the attention mechanism. Comprehensive experiments, using rigorous reconstruction evaluation metrics, demonstrate that Time-series ViT framework significantly outperforms baselines that use non-time-series MSI and SAR or time-series MSI without SAR, effectively enhancing MSI image reconstruction in cloud-covered regions.

Federated Learning (FL) is a collaborative machine learning paradigm which allows participants to collectively train a model while training data remains private. This paradigm is especially beneficial for sectors like finance, where data privacy, security and model performance are paramount. FL has been extensively studied in the years following its introduction, leading to, among others, better performing collaboration techniques, ways to defend against other clients trying to attack the model, and contribution assessment methods. An important element in for-profit Federated Learning is the development of incentive methods to determine the allocation and distribution of rewards for participants. While numerous methods for allocation have been proposed and thoroughly explored, distribution frameworks remain relatively understudied. In this paper, we propose a novel framework which introduces client-specific tokens as investment vehicles within the FL ecosystem. Our framework aims to address the limitations of existing incentive schemes by leveraging a decentralized finance (DeFi) platform and automated market makers (AMMs) to create a more flexible and scalable reward distribution system for participants, and a mechanism for third parties to invest in the federation learning process.

Mapping target crops before the harvest season for regions lacking crop-specific ground truth is critical for global food security. Utilizing multispectral remote sensing and domain adaptation methods, prior studies strive to produce precise crop maps in these regions (target domain) with the help of the crop-specific labelled remote sensing data from the source regions (source domain). However, existing approaches assume identical label spaces across those domains, a challenge often unmet in reality, necessitating a more adaptable solution. This paper introduces the Multiple Crop Mapping Generative Adversarial Neural Network (MultiCropGAN) model, comprising a generator, discriminator, and classifier. The generator transforms target domain data into the source domain, employing identity losses to retain the characteristics of the target data. The discriminator aims to distinguish them and shares the structure and weights with the classifier, which locates crops in the target domain using the generator’s output. This model’s novel capability lies in locating target crops within the target domain, overcoming differences in crop type label spaces between the target and source domains. In experiments, MultiCropGAN is benchmarked against various baseline methods. Notably, when facing differing label spaces, MultiCropGAN significantly outperforms other baseline methods. The Overall Accuracy is improved by about 10%.

Anomaly detection in financial transactions poses significant privacy challenges. This paper introduces a federated learning (FL) framework for Privacy-Preserving Behavioral Anomaly Detection using Graph Neural Networks (GNNs) on dynamic graphs to model cardholder transactions. We incorporate anonymization-based and noise-based privacy-preserving methods for feature engineering and a domain-specific negative sampling technique to train models without labeled data, making it suitable for real-world applications. Our results, benchmarked on synthetic and real-world datasets, show that deep learning-based outperform clustering-based methods, with F1-scores of 0.91 ± 0.02 and 0.87 ± 0.04, respectively. Additionally, using the anomaly score as a feature in fraud detection models yields a 1.76

Over the past years, Decentralized Finance (DeFi) protocols have suffered from several attacks. As a result, multiple solutions have been proposed to prevent such attacks. Most solutions rely on identifying malicious transactions before they are included in blocks. However, with the emergence of private pools, attackers can now conceal their exploit transactions from attack detection. This poses a significant challenge for existing security tools, which primarily rely on monitoring transactions in public mempools. To effectively address this challenge, it is crucial to develop proactive methods that predict malicious behavior before the actual attack transactions occur. In this work, we introduce a novel methodology to infer potential victims by analyzing the deployment bytecode of malicious smart contracts. Our idea leverages the fact that attackers typically split their attacks into two stages, a deployment stage, and an attack stage. This provides a small window to analyze the attacker’s deployment code and identify victims in a timely manner before the actual attack occurs. By analyzing a set of past DeFi attacks, this work demonstrates that the victim of an attack transaction can be identified with an accuracy of almost 70%.