{"slug": "dynamics-as-code-on-model-compression-via-dynamic-system", "title": "Dynamics as Code: On Model Compression via Dynamic System", "summary": "A new arXiv paper (2610.11115v1) proposes a generalized dynamic-system (DS) framework for neural network model compression that encodes high-dimensional weights as trajectory indices rather than pruning, quantizing, distilling, or low-rank decomposing them. The authors prove that under a Diophantine condition a finite trajectory of M = O(ε^-(d+ν)) states in an irrational winding forms an ε-net over the d-dimensional weight space, linking state resolution, decompression error, and compression ratio predictably, and unify four DS families: space-filling curves (Hilbert, Peano, Morton/Z-order, Snake), chaotic systems (Lorenz), congruential and pseudo-random generators (LCG, PCG), and low-discrepancy sequences (Halton). Experiments on ResNet-18 and Qwen2.5-1.5B/Qwen1.5-7B show competitive compression ratios without post-hoc retraining, with KD-tree and coordinate-template acceleration for large models and outlier identification to control error.", "body_md": "arXiv:2610.11115v1 Announce Type: new \nAbstract: The escalating size of pretrained neural networks has rendered model compression a prerequisite for deployment under stringent memory and compute constraints. With the irrational winding as an example, earlier work introduced a dynamic system (DS) paradigm that reconceptualizes compression as compact weight representation: high-dimensional parameters are encoded by the index of a trajectory produced by a dynamic system, from which the vector is recovered during decompression. This mechanism is fundamentally distinct from pruning, quantization, knowledge distillation, and low-rank decomposition. Along this direction, we prove that under a Diophantine condition, a finite trajectory of $M = O(\\epsilon^{-(d+\\nu)})$ states in the irrational winding constitutes an $\\epsilon$-net over the $d$-dimensional weight space, thereby linking state resolution, decompression error, and compression ratio in a predictable manner. Furthermore, we propose a generalized DS-based model compression framework by unifying four DS families---space-filling curves (Hilbert, Peano, Morton/Z-order, Snake), chaotic systems (Lorenz), congruential and pseudo-random generators (LCG, PCG), and low-discrepancy sequences (Halton). Also, we introduce the KD-tree and coordinate-template acceleration to scale to large models as well as outlier identification to control the error. Experiments on ResNet-18 and Qwen2.5-1.5B/Qwen1.5-7B validate that DS-based compression achieves competitive compression ratios without post-hoc retraining, with controllable decompression error and flexible state-space design, establishing it as a principled and practical compression approach.", "url": "https://wpnews.pro/news/dynamics-as-code-on-model-compression-via-dynamic-system", "canonical_source": "https://www.machinebrief.com/news/dynamics-as-code-on-model-compression-via-dynamic-system-x6j1", "published_at": "2026-10-09 04:00:00+00:00", "updated_at": "2026-10-09 05:46:50.001661+00:00", "lang": "en", "topics": ["machine-learning", "large-language-models", "ai-research", "ai-infrastructure"], "entities": ["arXiv", "ResNet-18", "Qwen2.5-1.5B", "Qwen1.5-7B", "Hilbert curve", "Lorenz system", "Halton sequence", "KD-tree"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/dynamics-as-code-on-model-compression-via-dynamic-system", "markdown": "https://wpnews.pro/news/dynamics-as-code-on-model-compression-via-dynamic-system.md", "text": "https://wpnews.pro/news/dynamics-as-code-on-model-compression-via-dynamic-system.txt", "jsonld": "https://wpnews.pro/news/dynamics-as-code-on-model-compression-via-dynamic-system.jsonld"}}