EKT-Router: Expert–Evidence Alignment for Embodied Vision–Language Models

Data release placeholder. Coming soon. No downloadable dataset is available in this repository. This is a documentation-only placeholder. It contains no examples from external benchmarks, training media, processed routing probes, or released data splits. No release date is promised. There is no dataset loading command at this stage.

Intended contents

Potential future artifacts include outcome-derived expert diagnostics, media-disjoint routing probes, paired route outcomes, typed packet records and evaluation manifests. The paper describes 1,020 fitting and 180 held-out routing probes; those records and split manifests were not included with the supplied release materials and are not present here. Intended schemas are documented in the source repository; its small synthetic smoke fixtures are software examples, not the experimental dataset.

Collection, processing, and quality

Exact source composition, provenance, split memberships, filtering, consent/rights review, normalization, and annotation procedures must accompany any actual release. The current card does not establish those properties. No new data collection, benchmark scoring, or human annotation was performed for this placeholder.

Access and licensing

No unified dataset license is assigned. Underlying benchmarks, media, retrieved text and annotations may have different terms; their original owners' terms remain applicable. Redistribution and per-source licensing must be checked before any future upload. The MIT code license does not grant rights to external data.

Limitations and reproducibility

The paper's 61.4, 74.7, and +1.75 refer to different evaluation protocols and are paper-reported results. They are not independently reproduced here. The missing weights, raw predictions and experiment logs prevent a reproduction claim.

Project and contact

• Project
• Code and synthetic schemas
• Model card
• Contact: rwlinno@gmail.com