| --- |
| language: |
| - en |
| pipeline_tag: image-text-to-text |
| tags: |
| - ekt-router |
| - embodied-question-answering |
| - lora |
| - coming-soon |
| --- |
| # EKT-Router: Expert–Evidence Alignment for Embodied Vision–Language Models |
|
|
| **Model release placeholder. Coming soon. No model weights or adapters are available to download.** |
| This repository contains a model card only. It is not a loadable Transformers model. |
| No release date is promised, and no third-party weights have been uploaded. |
|
|
| ## Model description |
|
|
| EKT-Router jointly selects one capability expert and optional knowledge/perception |
| under a bounded acquisition budget. The intended portfolio uses a shared |
| Qwen3-VL-8B backbone and three independent LoRA adapters: General (continued |
| pretraining), Spatial Reasoning, and Spatiotemporal. A frozen Qwen3-VL-2B descriptor |
| feeds the utility router. After one acquisition round, the Packet Verifier retains |
| valid evidence with positive predicted gain while keeping the expert fixed. The |
| expert generates once, using its original input if all evidence is rejected. |
|
|
| Knowin describes this project as its first internal Mixture-of-LoRA (MoL) solution; |
| this is an internal project milestone, not an academic priority claim. |
|
|
| ## Availability and reproducibility |
|
|
| The General, Spatial Reasoning and Spatiotemporal adapters, trained router, verifier, |
| and capability signatures are not released. The supplied project materials do not |
| include original predictions, complete experiment logs, immutable split/scorer |
| manifests, or weights sufficient to reproduce the paper measurements. |
|
|
| 61.4 (six embodied QA benchmarks), 74.7 (ten general VLM benchmarks), and +1.75 |
| (nine-benchmark evidence-null control) are paper-reported measurements under |
| separate protocols. They are not results reproduced by this public release. |
| CPU mock demos and unit tests are software checks, not model evaluation. |
|
|
| ## Intended use and limitations |
|
|
| Research on embodied vision-language question answering and expert/evidence routing. |
| Packet checks and fallback do not establish factual correctness, physical safety, |
| or suitability for autonomous robot deployment. Domain transfer, robustness, |
| and calibration require independent evaluation. |
|
|
| ## Training and licensing |
|
|
| See the paper and source repository for the reported methodology. Missing training |
| artifacts must not be inferred from this card. Weight licensing will be documented |
| per artifact if and when weights become available; the code's MIT license does not |
| license unreleased weights or third-party models. |
|
|
| ## Project and contact |
|
|
| - [Project](https://knowin-29.github.io/KnowinRouter.github.io/) |
| - [Code](https://github.com/Knowin-29/EKT-Router) |
| - [Data card](https://huggingface.co/datasets/Knowin-29/EKT-Router) |
| - Contact: rwlinno@gmail.com |
|
|
| Authors: Weilin Ruan, Yingjie Xu, Yangkai Wei, Bohan Zhou, Shuaibo Li, |
| Harold Haodong Chen, Kaiwen Zhou, Yinchuan Li, Ying-Cong Chen, James Cheng (corresponding author). |
|
|
