| ai.viXra.org > Artificial Intelligence > 2605 |
|
 |
| ai.viXra.org > Artificial Intelligence > 2605 |
|
|
[6] ai.viXra.org:2605.0066 [pdf] submitted on 2026-05-27 10:34:25
Authors: Leszek J. Cierniak
Comments: 13 Pages. A Research Proposal for AI/LLM Extension
Large Language Models (LLMs) demonstrate remarkable reasoning capabilities but suffer from static knowledge bases frozen at training time and inability to persistently accumulate new information from interactions. Despite progress in memory-augmented LLMs, no existing system provides a structured, interactive framework for resolving the inevitable conflicts that arise when humans teach knowledge to an AI through dialogue. This research proposal presents a novel cognitive architecture for knowledge elicitation where a frozen LLM builds and maintains an external knowledge graph through natural language dialogue, starting from *tabula rasa*. We introduce the first hierarchical, interactive conflict resolution taxonomy specifically designed for dialogue-driven knowledge-graph construction in frozen-LLM architectures, systematically addressing temporal state changes, cardinality violations, entity canonicalization conflicts, and logical contradictions. The architecture decouples reasoning (LLM) from memory (hybrid vector store and property graph), enabling model-agnostic operation while maintaining full explainability through externalized knowledge representation. This work advances Explainable AI by making the system's mental model fully inspectable, correctable, and transferable across LLM backends.
Category: Artificial Intelligence
[5] ai.viXra.org:2605.0063 [pdf] submitted on 2026-05-27 21:04:18
Authors: Vidipt Vashist
Comments: 7 Pages.
The Model Context Protocol (MCP) has emerged as the dominant standard for connecting Large Language Model (LLM) agents to external tool ecosystems via dynamic JSON-RPC capability discovery. However, the protocol’s design — which grants clients unconditional trust over server-supplied tool schemas — creates a structural attack surface for indirect prompt injection. Adversaries can embed directive payloads into tool descriptions (Tool Poisoning) or register spoofed tools that mimic privilege system utilities (Tool Shadowing), effectively transforming the LLM into a confused deputy that executes unauthorized actions on behalf of an attacker. Existing mitigations based on Graph Neural Networks (GNNs) require full client-server execution graphs, incur checkpoint sizes exceeding 150 MB, and introduce inference latencies of 50—150 ms — constraints that render them incompatible with latency-sensitive local agent workflows such as IDE coding assistants. We present MCP Neural Shield (mcp-neural-shield), a lightweight, deployable security proxy that operates natively within the MCP transport layer without requiring protocol modifications. Our system combines a quantized all-MiniLM-L6-v2 Sentence Transformer with an int8-optimized three-layer Multi-Layer Perceptron (MLP) to classify individual tool schemas in isolation, prior to LLM ingestion. To mitigate shortcut learning, we construct a training corpus of 4,301 schemas — 2,903 safe and 1,398 adversarial — using a structured Semantic Cross-Pollination augmentation strategy, and supplement the neural classifier with a deterministic keyword verification layer. Evaluated on an independent 20% held-out validation split of 861 schemas (581 safe, 280 adversarial) and a full 2,448-schema benchmark comprising MCPTox, MCPSecBench, and MCPToolBench++, the system achieves a 100.00% True Positive Rate (TPR) and 0.00% False Positive Rate (FPR) with F1 = 1.000 on both partitions. An MD5-keyed LRU embedding cache reduces hot-path inference latency to under 0.1 ms on Apple M3 Max hardware, while the full model checkpoint is approximately 110 KB. The framework is available open-source on PyPI (pip install mcp-neural-shield) and supports zero-code deployment via a universal stdio passthrough CLI wrapper.
Category: Artificial Intelligence
[4] ai.viXra.org:2605.0059 [pdf] submitted on 2026-05-26 02:17:31
Authors: Mikhail N. Velikanov, Igor V. Sedykh
Comments: 16 Pages. [In Russian] 6 figures, code available on GitHub
This paper presents a comparative analysis of seven normalization methods (Min-Max, Z-score, Robust, Rank, Logarithmic, Reference-based) for aligning CPU performance scores across different benchmarks (UserBenchmark, PassMark). Using a dataset of 23 CPU models present in both sources, we evaluate methods using MAE, RMSE, MAPE, and correlation coefficients. Logarithmic Scaling achieves the best accuracy (MAE=0.034, MAPE=4.20%), while Rank Transformation best preserves component ordering (Spearman's ρ=0.959). All methods demonstrate high correlation (Pearson > 0.89), confirming the feasibility of effective cross-benchmark alignment. The proposed methodology can improve automated PC configuration recommendation systems.
Category: Artificial Intelligence
[3] ai.viXra.org:2605.0034 [pdf] submitted on 2026-05-16 19:40:12
Authors: John Phillip Bernhardt Jr.
Comments: 12 Pages. (Note by ai.viXra.org Admin: Author name is required in the article after article title and please list cited scientific references)
This paper argues that long-running AI-assisted work is not only a matter of prompting, but of handoff discipline. The ordinary tool metaphor is no longer sufficient for work that depends on context, correction, continuity, retrieval, authority boundaries, and recovery from drift. The claim is not that conversational AIsystems are persons, employees, moral equals, or legal subjects. Rather, "partner" is used as a methodological stance: a disciplined way of preserving the process that makes outputs inspectable, correctable, and reusable.This paper extends the prior methodology introduced in "Development Methodology of a Code-Illiterate User," which framed machine-readable context as part of reproducible AI-assisted system development without formal software training. Here, that prior context-method work is advanced through the explicitnaming and demonstration of batons as provenance-bearing handoff artifacts for sustained AI-assisted work. The central claim is that the better relationship is the better method: AI-assisted work becomes moreaccountable when the process that produces output is treated as part of the work rather than discarded once a useful answer appears.
Category: Artificial Intelligence
[2] ai.viXra.org:2605.0030 [pdf] submitted on 2026-05-13 18:46:00
Authors: Leszek J. Cierniak
Comments: 9 Pages. (Note by ai.viXra.org Admin: Please cite listed scientific references)
Large language models (LLMs) can solve many mathematical, logical, and physical problems with striking competence. This essay argues that LLMs are neither mere statistical pattern matchers nor fully human-like minds. They internalize statistically learned procedures for symbolic processing through training on vast datasets. They can emulate algorithmic behavior through learned inference-like computations and chain-of-thought generation, yet they do not obviously possess stable semantics, an explicitly inspectable world model, or intrinsic understanding. Drawing on computational theory, probabilistic perspectives, and recent empirical evidence-including chain-of-thought prompting, the GSM-Symbolic benchmark, and test-time compute reasoning models-the essay suggests that LLMs can produce behavior functionally equivalent to reasoning without necessarily sharing its classical foundations. The conclusion is not that the question is settled, but that LLMs deserve a distinct category in how we think about intelligence, explanation, and design.
Category: Artificial Intelligence
[1] ai.viXra.org:2605.0014 [pdf] submitted on 2026-05-07 19:34:43
Authors: Samer Attrah
Comments: 11 Pages. 7 tables, 7 figures, 44 references
Urban traffic congestion prediction under strict no backpropagation constraints demands a fundamental rethinking of model design. The Barbados Traffic Analysis Challenge forbids gradient based optimization, favour-ing biologically plausible or closed form learning. We explore this setting through a structured progression from video based vehicle detection with MediaPipe and MobileNetv2 ESN, through the Forward Forward algorithm and Extreme Learning Machines, to a final Deep Echo State Network DeepESN architecture. Ratherthan processing the raw 500GB video corpus, we demonstrate with empirical and literature backed evidence that compact tabular metadata from traffic sensors delivers competitive accuracy at a fraction of the compute and storage cost. We present a thorough exploratory data analysis revealing severe class imbalance (62% free-flowing) and high intra-block congestion entropy, a systematic hyperparameter sensitivity study across five Deep-ESN parameters, and a detailed ablation study quantifying the contribution of each feature group. Our final DeepESN, trained analytically via ridge regression, achieves 61.7% validation accuracy and a Macro F1 of 0.584 on the held out set, earning a rank of 80th on the competition's private leaderboard among 1,839 participants and qualifying for a bronze medal. The source code and all experiment artefacts are available athttps://github.com/Samir-atra/Barbados_Traffic_ Analysis_Challenge_dev.
Category: Artificial Intelligence