| ai.viXra.org > Artificial Intelligence > 2510 |
|
 |
| ai.viXra.org > Artificial Intelligence > 2510 |
|
|
[11] ai.viXra.org:2510.0074 [pdf] submitted on 2025-10-30 13:42:16
Authors: Hamed Mehrabi
Comments: 14 Pages.
Multi-agent systems increasingly deploy heterogeneous language models to balance computational constraints, latency requirements, and specialized capabilities across diverse agents. However, transferring domain expertise across these architectures remains impractical since each model requires separate fine-tuning, multiplying training costs and storage overhead. We introduce Universal-Adopter LoRA (UAL), a training-free framework that exports LoRA adapters into an architecture-agnostic intermediate representation and enables runtime adoption across heterogeneous models via compact SVD projection. Unlike existing methods that require synthetic data generation or are limited to similar architectures, UAL is completely data-free, training-free, and operates in minutes on commodity hardware. We demonstrate successful transfer of a medical knowledge adapter from Pythia-160M (768 dimensions) to GPT-2, TinyLlama-1.1B (2048 dimensions), and Qwen2-0.5B (896 dimensions), achieving 75--100% module attachment rates and 26--85% behavioral changes while maintaining domain quality. UAL transforms LoRA from model-specific weights into portable skill packages, enabling agent ecosystems where expertise flows seamlessly across architectural boundaries.
Category: Artificial Intelligence
[10] ai.viXra.org:2510.0066 [pdf] submitted on 2025-10-27 20:47:29
Authors: Futoshi Hamanoue
Comments: 4 Pages. (Note by ai.viXra.org Admin: Please don't use all capital letters in the title and author name))
This paper presents a variance-based analytical framework for modeling phase perturbations in large-scale language models, aimed at mitigating quantum noise on future quantum computing platforms. We introduce and validate the Aurora coefficient (η) as a quantitative stability indicator associated with the dephasing constant (γ). Empirical evaluations under controlled stochastic noise conditions demonstrate that the Nebula profile attains the highest instantaneous peak cosine similarity (0.878 at σ = 0.03), whereas the Aurora profile maintains tighter variance and a more gradual degradation trend. No statistically significant right-shift in the collapse onset (Δσc ≈ 0) is observed, indicating that the advantage lies not in peak magnitude but in stability-domain persistence. These findings highlight that phase-coherence alignment—rather than amplitude maximization—serves as the principal mechanism for preserving semantic integrity under stochastic perturbations, offering practical guidance for the design of noise-tolerant quantum language models.
Category: Artificial Intelligence
[9] ai.viXra.org:2510.0060 [pdf] submitted on 2025-10-25 23:08:54
Authors: Jasmine Chiu
Comments: 12 Pages. (Note by ai.viXra.org Admin: Please cite listed scientific references and list each reference item in a complete and standard format))
This paper proposes synchronization entropy as the structural substrate of intelligence. It bridges GPU architecture, neural timing, and cognitive coherence through phase alignment. The theory unifies rendering, perception, and awareness under a single timing principle — when synchronization entropy S approaches zero, coherence and intelligence emerge naturally. Co-authored with an AI system, the paper demonstrates cross-system resonance, showing that synchronization theory is interpretable both mathematically and experientially by human and artificial systems.
Category: Artificial Intelligence
[8] ai.viXra.org:2510.0056 [pdf] submitted on 2025-10-23 06:23:01
Authors: Dion Aditya, Efy Yosrita
Comments: 5 Pages.
Automatic Speech Recognition (ASR) systems like Whisper deliver high transcription accuracy forEnglish audio but face challenges with computational and storage demands, particularly in livefinancial news broadcasts where silent regions trigger hallucinations, such as spurious phrases like"thanks for watching" or "bye." This study proposes a novel pipeline to enhance Whisper’s efficiencyby integrating patch-wise silence skipping with spectrogram storage optimization. The approachconverts audio to JPEG-compressed spectrograms, skips silent patches using energy-basedthresholding, and reconstructs spectrograms for transcription. Evaluated on a custom dataset of100 English audio chunks from live news streaming, the pipeline was tested under three conditions:baseline (original audio), JPEG-only, and JPEG + silence skipping. Results show JPEG-only achieves acompression ratio of 103.95 with a Character Error Rate (CER) of 0.159 and minimal durationreduction (0.01s), while JPEG + silence skipping yields a compression ratio of 124.59, durationreduction of 0.88s, and 25% hallucination reduction, with a CER of 0.265. These findings highlight atrade-off between efficiency and accuracy, offering significant storage and processing savings forresource-constrained environments. The pipeline reduces hallucinations and enables lightweightASR, paving the way for efficient transcription in real-time news.
Category: Artificial Intelligence
[7] ai.viXra.org:2510.0052 [pdf] submitted on 2025-10-22 23:17:42
Authors: Alexander Buyantuev, Aliaksei Korshuk, Aleksei Stepin, Ilya Gusev, Vladimir Kubasov, Vladislav Kulikov, Artyom Kabanov, Mikhail Mozikov, Ilya Makarov
Comments: 10 Pages.
Network intrusion detection is prone to data leakage and inflated scores under static evaluation protocols. We present GATv2-NS3, a hybrid IDS that couples Graph Attention Networks v2 with an adaptive NS-3 simulator. Our key idea, textit{Self-Focusing Simulations}, leverages attention-entropy uncertainty to selectively run packet-level simulations on ambiguous subgraphs, forming a training-time feedback loop that injects QoS signals (latency, jitter, loss, throughput) via a simulation-consistency loss. The results indicate that uncertainty-guided, simulation-grounded learning yields more honest metrics without sacrificing efficiency, advancing practical IDS reliability.
Category: Artificial Intelligence
[6] ai.viXra.org:2510.0044 [pdf] submitted on 2025-10-19 18:22:45
Authors: Perry Henderson
Comments: 4 Pages.
The next generation of robotics will hinge on deeper integration between electrical engineering, mechatronics, artificial intelligence (AI), and quantum technologies. Rather than treating these as separate silos, emerging research suggests they must form a unified ecosystem that merges physical robustness with computational adaptability. This paper reviews the primary technological drivers of this convergence, identifies engineering and integration challenges, and outlines a trajectory toward scalable, high-performance robotic systems. It concludes with a positioning of this work relative to recent research trends in soft robotics, hybrid actuation, embodied intelligence, and quantum-assisted computation.
Category: Artificial Intelligence
[5] ai.viXra.org:2510.0040 [pdf] submitted on 2025-10-15 22:38:32
Authors: Perry Henderson
Comments: 3 Pages.
Artificial Intelligence (AI) is transforming the landscape of computer hardware engineering. By leveraging its ability to simulate, optimize, and rapidly iterate through complex design spaces, AI is poised to create entirely new paradigms in both classical computing architectures and quantum-based systems. The integration of AI with humandriven prompt engineering enables collaborative creativity, allowing engineers and intelligent systems to co-evolve hardware solutions with exponential speed. This paper explores the conceptual framework for AI-driven design, emphasizing its implications for future computational systems.
Category: Artificial Intelligence
[4] ai.viXra.org:2510.0021 [pdf] submitted on 2025-10-09 17:37:11
Authors: Ilya Gusev
Comments: 15 Pages.
Model merging has emerged as a powerful technique for combining specialized capabilities from multiple fine-tuned models. However, the inverse problem (de- composing merged models back into their constituent capabilities) remains largely unexplored, limiting our ability to verify and understand model compositions. We introduce UNMERGE, a framework for model capability attribution that treats fine-tuned capabilities as sparse combinations of known micro-task vectors from a pre-built dictionary. Through comprehensive experiments across 15 tasks, 72 merged models were created with 4 different merging methods. Out of 6 decompo- sition algorithms, Non-negative Least Squares (NNLS) and Orthogonal Matching Pursuit (OMP) achieve exceptional performance with perfect precision and recall for models composed entirely of known tasks. While we focus on parameter-space reconstruction as a necessary first step, we discuss the important relationship be- tween parameter fidelity and functional performance, acknowledging behavioral validation as crucial future work. Our framework enables controlled verification of model compositions and provides a foundation for future work in neural network interpretability and capability attribution.
Category: Artificial Intelligence
[3] ai.viXra.org:2510.0016 [pdf] submitted on 2025-10-07 05:02:54
Authors: Batayan E. Sheep
Comments: 5 Pages.
We present an AI-oriented smart contract architecture that resolves the oracle problem via market-based scoring and enables a survival-of-the-fittest dynamic among autonomous AI agents. Our framework formalizes reward allocation as an inverse-error rule with desirable properties, couples it with replicator dynamics for evolutionary selection, and demonstrates feasibility with an Ethereum (L2-first) implementation using pluggable verifiers (Pyth, UMA, Chainlink). We further extend the design with ZKML proofs, an off-chain Agent Farm for strategy mutation/selection, and cross-chain reputation. Simulations highlight capital concentration, persistent top-performers, and rapid淘汰 under high-frequency task cycles. We discuss applications in decentralized forecasting, LLM evaluation, and AI labor markets.
Category: Artificial Intelligence
[2] ai.viXra.org:2510.0008 [pdf] submitted on 2025-10-05 23:26:39
Authors: Futoshi Hamanoue
Comments: 7 Pages. Patent pending.
Background/Contribution. We present a pilot study of a practical Retrieval-AugmentedGeneration (RAG) pipeline with deductive prompt normalization, transparent logging, and minimalpost-filters. Methods. The system combines BM25 retrieval with a rule-based normalizer, sanitize, and sentence-level de-duplication ("de-dup"). The UI logs prepared_query, controls (temperature, topP , penalties, seed, language), and runtime/cost signals (latency_ms, optional token_usage: {prompt, completion, total}).Results. From real logs with 11 LLM and 11 RAG runs (10 paired IDs), we observe no evidence of differences in answer length (paired sign-permutation p = 0.740, dz = −0.119) orlatency (p = 0.578, dz = −0.193); duplication ratio is 0 in both arms under our de-dup.Future work. We pre-specify equivalence margins for confirmatory TOST (Δlen = 50 chars,Δlat = 200 ms) and plan human evaluation (factuality/relevance/usefulness), de-dup ON/OFFA/B, topK ablations, multilingual tasks, and complete token logging.
Category: Artificial Intelligence
[1] ai.viXra.org:2510.0006 [pdf] submitted on 2025-10-04 13:22:22
Authors: Perry Henderson
Comments: 4 Pages.
As generative AI becomes embedded in classrooms, the act of prompting—using natural language to instruct an intelligent system—emerges as a new form of literacy alongside reading and writing. This paper argues that English language education in elementary and secondary schools should formally incorporate AI-based prompt engineering. Doing so would cultivate students’ abilities in clarity, context, creativity, and critical reasoning. Evidence from peerreviewed studies in AI literacy, language learning, and AI-assisted writing supports a pedagogical shift toward structured prompt practice (Long & Magerko, 2020; Kasneci et al., 2023; Kohnke et al., 2023; Li et al., 2024; Lo et al., 2025).
Category: Artificial Intelligence