Vol 13 No 1 (June 2026): OPEN

Title: Capability-Specific Degradation Patterns in Quantized Small Language Models
Author(s): Emil Rahimov
Abstract: Post-training quantization is the default method for deploying small language models (one to four billion parameters) on consumer and edge hardware, yet its effect is usually summarized by a single aggregate accuracy score that can conceal severe failures in individual capabilities. This paper presents a capability-level analysis of four-bit quantization for seven open instruction-tuned small language models drawn from five architecture families: Qwen2.5, Llama-3.2, Gemma-2, Phi-3.5, and SmolLM2. Each model is evaluated at sixteen-bit floating point and at four-bit precision across six capabilities, namely factual knowledge, commonsense reasoning, mathematical reasoning, multilingual mathematical reasoning, code generation, and instruction following, yielding eighty-four controlled evaluations. Instead of reporting only mean accuracy, we construct a per-capability degradation map and test, using Kendall's rank correlation, whether capabilities degrade in a consistent order across architectures. The results show that degradation is strongly capability-specific: multilingual mathematical reasoning and code generation are the most fragile capabilities, with relative losses of up to fifty-seven percent, whereas commonsense reasoning is almost entirely preserved. However, the ordering of degradation is only weakly consistent across architectures, with a mean Kendall's tau of 0.29, indicating that the most and least fragile capabilities are shared but the overall ranking is architecture-dependent. Smaller models degrade more in magnitude. We conclude that small-model quantization should be evaluated per capability rather than in aggregate.
DOI: 10.36079/lamintang.ijai-01301.1050

Title: A Standard Deep Learning-Based Model That Integrates R3D-18 With RNN for Clip Classification and Transition Point Detection in Basketball Video Footages
Author(s): Paul Aliga Paul, Joshua Ojo Nehinbe, Kingsley Eghonghon Ukhurebor
Abstract: Research on basketball video clip analysis is rapidly emerging as a vital tool for coaching, player development, scouting, recruitment of coaches and players. Such analysis enhances post-game review, tactical planning, practice design, in-game adjustments, visual feedback and decision-making. However, many existing models lack robust architectures for transition detection, limiting their effectiveness. Additionally, high costs, limited customization, internet dependency and complex interfaces restrict widespread adoption of most models across professional and grassroots levels. To address these challenges, this paper implements a deep learning model with Python language. The model integrates Residual 3D Network with 18 layers (R3D-18) to capture spatial-temporal features, Convolutional neural networks (CNNs) to perform feature extraction and Recurrent Neural Network (RNN) to accomplish temporal modeling of offline basketball videos. Using entropy thresholding and top-k (i.e. top-10) accuracy metrics on four basketball datasets, the model performs clip-wise and video-level action classification with exceptional confidence, achieving 99.92% and 99.52% certainty that all the evaluative footages depict basketball videos. The model further detected no transitions in basketball_video1 and basketball_video3 while notable transitions appeared at clips 8 and 7 in basketball_video2 and basketball_video4, respectively corresponding to spikes in uncertainty likely caused by activity changes like switching from basketball to running. Average entropy scores of 4.787, 3.5341, 3.7912 and 3.1976 across datasets were reported, reflecting potentially elusive variations and uncertainties within clip-level predictions due to nuanced activity shifts despite strong overall classification confidence.
DOI: 10.36079/lamintang.ijai-01301.883