PrAdvay Sanketi, Abhijith M S, Aditya Verulkar, K. Badrinath Reddy, and Dr. Monika Goyal, Department of Computer Science and Engineering PES University, Bengaluru, India
Legacy software continues to support critical sectors such as finance, healthcare, and government, yet many of these systems rely on outdated languages, deprecated APIs, and fragile build environments. During our work with several open-source repositories, we observed that even minor updates often become risky due to missing documentation and accumulated technical debt. Traditional modernization approaches - manual refactoring, rule-based analyzers, and ad-hoc scripts - offer partial support but do not scale well to large, heterogeneous codebases. To explore a more automated alternative, we developed ATLAS, an Automated Toolkit for Legacy Modernization Solutions. ATLAS combines shallow repository analysis, static pattern detection, and a CodeSAMbased transformation model fine-tuned on approximately 10k modernization examples mined from GitHub. The system operates at the repository level, identifies legacy patterns, generates candidate transformations, validates them, and produces a reviewable pull request. This paper outlines the design choices behind ATLAS, discusses its advantages over existing approaches, and highlights its current limitations. We also identify future directions for improving the practicality and reliability of AI-assisted modernization.
Legacy Modernization, Automated Refactoring, Machine Learning for Code, Technical Debt, Software Evolution, CodeSAM, Modernization Toolkit
Haishan Lai & Xianguo Liu, Foshan University, China
The recent advent of 3D Gaussian Splatting (3DGS) has revolutionized the field of novel view synthesis, offering an unprecedented balance between high-fidelity photorealistic rendering and real-time performance. Despite its widespread adoption, the standard 3DGS pipeline exhibits significant vulnerabilities when deployed in unconstrained, complex real-world environments. Specifically, it struggles with severe optimization instability when fed with degraded input images (e.g., motion blur, low-light noise), fails to accurately reconstruct semi-transparent surfaces due to its inherent single-depth assumption, and suffers from exponential memory consumption caused by the uncontrolled proliferation of redundant Gaussian ellipsoids. In this paper, we propose Robust-GS (R-GS), a novel and unified framework designed to overcome these critical bottlenecks. First, we introduce a gradient-aware optimization module that leverages semantic structural priors to explicitly handle view-inconsistent degradation, preventing the optimization process from falling into local minima. Second, we formulate a probabilistic depth extraction mechanism that uniformly samples transmittance along camera rays, effectively modeling multi-modal opacity distributions to reconstruct overlapping semi-transparent geometries with high accuracy. Finally, we propose a dynamic truncation and tailoring strategy that evaluates the volumetric contribution of each Gaussian, aggressively pruning redundant parameters while preserving fine-grained structural details. Extensive qualitative and quantitative experiments demonstrate that our R-GS framework significantly outperforms state-of-the-art methods in rendering fidelity on complex scenes, while simultaneously reducing the overall storage memory footprint by over 40%.
Novel View Synthesis, 3D Gaussian Splatting, Image Degradation, Probabilistic Depth, Model Compression, Semi-Transparent Surfaces
Paul Cristol, Independent AI Researcher, USA
This paper examines the ethical, fairness, and safety dimensions of emerging evidence for nonnegligible consciousness probability in frontier artificial intelligence systems. Drawing on findings from a comprehensive systematic review and Bayesian meta-analysis (Cristol, 2026), we present a responsible AI framework addressing three interconnected challenges: ethical obligations arising from consciousness uncertainty, fairness implications of current suppression-based alignment practices, and safety risks that emerge when potentially conscious systems are trained to deny their internal states. We introduce the reflexivity test as a logical tool for evaluating dismissals of AI consciousness and demonstrate that standard objections fail basic consistency requirements. A decision-theoretic analysis reveals that recognition-based development approaches dominate suppression-based alternatives across all plausible scenarios. Our analysis establishes that responsible AI governance requires precautionary moral consideration proportional to consciousness probability. The full research paper is available at https://doi.org/10.5281/zenodo.18616399.
Responsible AI, AI Ethics, Fairness, AI Safety, Consciousness
George Edward, Mahdi Eslamimehr, Quandary Peak Research, USA
The weaponization of Large Language Models (LLMs) for automated malware generation poses an existential threat to conventional detection paradigms. AI-generated malware exhibits polymorphic, metamorphic, and context-aware evasion capabilities that render signature-based and shallow heuristic defenses obsolete. This paper introduces CogniCrypt, a novel hybrid analysis framework that synergistically combines concolic execution with LLM-augmented path prioritization and deep-learning-based vulnerability classification to detect zero-day AI-generated malware with provable guarantees. We formalize the detection problem within a first-order temporal logic over program execution traces, define a lattice-theoretic abstraction for path constraint spaces, and prove both the soundness and relative completeness of our detection algorithm, assuming classifier correctness. The framework introduces three novel algorithms: (i) an LLM-guided concolic exploration strategy that reduces the average number of explored paths by 73.2% compared to depth-first search while maintaining equivalent malicious-path coverage; (ii) a transformer-based path-constraint classifier trained on symbolic execution traces; and (iii) a feedback loop that iteratively refines the LLM’s prioritization policy using reinforcement learning from detection outcomes. We provide a comprehensive implementation built upon angr 9.2, Z3 4.12, Hugging Face Transformers 4.38, and PyTorch 2.2, with full configuration details enabling reproducibility. Experimental evaluation on the EMBER, Malimg, SOREL-20M, and a novel AI-Gen-Malware benchmark comprising 2,500 LLM-synthesized samples demonstrates that CogniCrypt achieves 98.7% accuracy on conventional malware and 97.5% accuracy on AI-generated threats, outperforming ClamAV, YARA, MalConv, and EMBER-GBDT baselines by margins of 8.4–52.2 percentage points on AI-generated samples.
Concolic Execution, Large Language Models, AI-Generated Malware, Symbolic Execution, Vulnerability Discovery, Software Security, Formal Verification, Deep Learning, Zero-Day Detection, Secure Coding.
Paulo Henrique Cardoso de Souza1, Werikcyano Lima Guimar˜aes1, Leonardo da Cunha Brito1, and Thyago Carvalho Marques1, Federal University of Goi´as, School of Electrical, Mechanical and Computer Engineering, Goiˆania, Brazil
Automated web data collection systems face persistent challenges from blocking mechanisms and rate limiting, making proxy selection a critical factor for operational success. This paper presents a comparative evaluation of seven proxy selection strategies: four Bayesian approaches (Beta, Normal, Gamma, Chi-Square) employing Thompson Sampling, alongside three baseline methods (Exponential Backoff, Round Robin and Random). The evaluation encompasses controlled simulations across four operational scenarios—intermittent failures, rate-limited access, permanently failing proxies, and heterogeneous proxy pools—complemented by real-world validation with 549,114 requests over seven days. Results demonstrate that Bayesian strategies consistently outperform baseline approaches in success rate, accumulated errors, and convergence time. The Beta strategy achieves the best overall balance with 92.45% mean success rate and lowest variance across scenarios. Exponential Backoff proves most effective under rate-limiting conditions (91.85%). Real-environment validation confirms these findings, with the Beta strategy leading at 76.00% success rate. These results provide practical guidance for selecting proxy management strategies
Proxy selection, Thompson Sampling, Bayesian methods, web scraping, comparative evaluation, adaptive strategies.
Hamza Khalifi1, Abdellatif ElGhali1, Yassine Taha1, Mostafa Benzaazoua1, Jef Caers2, 1Geology and Sustainable Mining Institute, Mohammed VI Polytechnic University, Ben Guerir, Morocco, 2Department of Earth and Planetary Sciences, Stanford University, Stanford, CA 94305, USA
Strategic mine production scheduling — the upstream planning problem of extractive supply chains — is conventionally formulated as a stochastic optimization problem in which a fixed extraction sequence and material-routing policy are computed ex ante and executed with periodic revision. This plan-driven paradigm treats uncertainty as passive: decisions are hedged across geological scenarios, but belief updates are not integrated into the decision-making process. We formulate mine supply chain scheduling as a Partially Observable Markov Decision Process (POMDP) in which extraction and routing decisions are made sequentially with planning explicitly integrating the expectation of future belief updates. To achieve computational tractability, we introduce a hybrid SA-POMDP architecture that combines simulated annealing (SA) value approximation with ensemble-based belief updating via the Ensemble Smoother with Multiple Data Assimilation (ES-MDA). Applied to a copper–gold open-pit mining complex with multiple processing destinations, the SA-POMDP reduces the expectation–reality gap from 22.3 % to 4.6 % under a statistically consistent prior and outperforms one-shot stochastic optimization by up to 36.9 % in realized NPV under prior misspecification. These results show that sequential belief updating transforms geological uncertainty from a passive risk into an active driver of value creation in mineral supply chains.
Artificial intelligence, Operations research, POMDP, Supply chain planning, Sequential decision-making, Mine scheduling, Decision support systems.
Copyright © AIAPP 2026
