vla-slmnegi: Integrating SigLIP Visual Patches with a 45M Parameter Decoder Backbone for Low-Latency Vision-Language-Action Robotics Control
vla-slmnegi is a lightweight, multimodal Vision-Language-Action (VLA) model developed by ArchiGen AI Labs Pvt Ltd. It connects a pre-trained SigLIP-base/16 vision encoder to the slmnegi 45.2M parameter Llama-decoder backbone through a 2-layer MLP projection alignment layer.
แ ๏ธ Read the Full Research Paper: Download vla_slmnegi_paper.pdf
Authors & Affiliation
- Lead Researcher: Manish Negi
- Organization: ArchiGen AI Labs Pvt Ltd
- Contact: manishnegi2032@gmail.com | manish@archgen.in | manish@vyvsai.com
Abstract
Vision-Language-Action (VLA) models unify perception, instruction comprehension, and robotic end-effector execution into a single autoregressive sequence generation framework. vla-slmnegi is an ultra-compact multimodal architecture engineered specifically for low-overhead multimodal pre-training and real-time inference on edge robotic platforms (<= 15 GB VRAM / NVIDIA T4 compatible).
System Architecture
| Component / Parameter | Value / Configuration |
|---|---|
| Vision Encoder | SigLIP-base/16 (768-dim, Frozen) |
| Projector Architecture | 2-Layer MLP ($768 \rightarrow 512 \rightarrow 512$) |
| Language Backbone | slmnegi (512-dim, 8 Layers, 8 Query / 2 KV Heads) |
| Visual Patch Tokens ($N_v$) | 196 tokens ($14 \times 14$ grid) |
| Precision Mode | FP16 Mixed Precision |
| Peak GPU VRAM Usage | ~5.2 GB (NVIDIA T4 Compatible) |
Key Theoretical Contributions
- Multimodal Projection Alignment: Maps 196 visual patch tokens ($768$-dim) into
slmnegi's $512$-dim hidden embedding space via a GELU-activated 2-layer MLP. - Target Label Masking Strategy: Prepends a $-100$ (
ignore_index) tensor over the 196 visual token positions in target sequences, ensuring cross-entropy loss is evaluated strictly over text instructions and action tokens to avoid dimension mismatches and gradient instability.
Citation
@article{negi2026vlaslmnegi,
title={vla-slmnegi: Integrating SigLIP Visual Patches with a 45M Parameter Decoder Backbone for Low-Latency Vision-Language-Action Robotics Control},
author={Negi, Manish},
journal={ArchiGen AI Labs Pvt Ltd Technical Report},
year={2026}
}