ZeGA: Zero-Gradient Gated Adaptation via Weight-Space Synapse Cartridges
Year: 2026
Type: preprint
Abstract: ZeGA (Zero-Gradient Gated Adaptation) compiles input-selective residual branches into appended Transformer MLP channels without gradient optimization. Building on Weight-Space Programming, we evaluate reversible sub-megabyte .safetensors Synapse Cartridges (aphygo.synapse_cartridge.v1) across language, vision-language, and SDXL image diffusion models, attached directly to native PyTorch nn.Linear layers via a 272-line zero-wrapper loader while leaving 100% of preexisting base channel weights untouched. We establish closed-form gate-value compilations for gated MLPs (Family I: SwiGLU, GeGLU, ReGLU) and prove exact 2-row stationary-minimum and 4-row fixed-center symmetric quartet polarization identities (Theorem 3) that synthesize bilinear gate-value products inside standard non-gated 2-layer MLPs (Family II: GELU, QuickGELU, SiLU) without a gate_proj matrix:1. Language (Ministral-3-8B-Instruct-2512-BF16, SwiGLU, 960.95 KiB): Improves SwedMaster-300 from 73.67% to 84.00% (+10.33 pp, p = 1.64e-6), SwedFreeGen-60 from 70.00% to 95.00% (p = 2.75e-4), and ScaLA-sv (n=1,000) from 71.30% to 75.40% (p = 0.00864) with zero collateral loss on Belebele-SV/EN/FR, Python-20, Math-20, and Anchor-40, and 100% SHA-256 base-weight restoration (8.40 ms hot-swap).2. Vision-Language (CLIP-ViT-B/32, QuickGELU, 336.80 KiB): Compiled in 26.6 s without gradient descent, our 4-row quartet raises Winoground-style Pair Group Score on CompBind-240 from 0.83% to 25.83% (p = 1.54e-8) and ClauseSwap-180 from 2.22% to 34.44% (p = 2.98e-8) with zero caption/template overlap while preserving CIFAR-100 and CelebA embeddings.3. Generative Image Diffusion (Juggernaut XL / SDXL, GEGLU, 906.84 KiB): Spatiotemporal patch-gated cartridges selectively transform target vessels across multi-object scenes while preserving nearby bystanders and achieving 32.31 dB full-image PSNR (0.9495 SSIM) on target-free anchors. Code & Benchmarks: https://github.com/Aphygo/zega-synapse-cartridgesCartridges (Hugging Face): https://huggingface.co/Aphygo/zega-ministral3-8b-swedish-v1 and https://huggingface.co/Aphygo/zega-clip-vit-b32-polarization-v1
Author: Thulin, Johan
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225867795
