Open work: MetroLLM-Bench published (arXiv 2609.10016), dataset link

README.md

@@ -38,7 +38,7 @@ Sovereignty here is technical, not geographic. It is about who controls the keys
 
 ## Open work
 
-MetroLLM-Bench is one example. It is a benchmark and a set of open-weight students for running a transit kiosk from a prose prompt. A small local language model handles the task offline on commodity hardware, and the 2.6 GB student matches a frontier cloud API.
+MetroLLM-Bench: a benchmark for running a transit kiosk from a prose prompt, with open-weight students that handle the task offline on commodity hardware. Published as [arXiv:2609.10016](https://huggingface.co/papers/2609.10016) (September 2026). The 2.6 GB student matches GPT-5.4 at maximum reasoning effort on the deterministic tier and exceeds both GPT-5.6 tiers.
 
 | Model | Size | Runs on |
 |-------|-----:|---------|
@@ -47,7 +47,7 @@ MetroLLM-Bench is one example. It is a benchmark and a set of open-weight studen
 | [Qwen3.5-9B-metro-v24](https://huggingface.co/continker/Qwen3.5-9B-metro-v24) | 5.3 GB | a 24 GB machine |
 | [Qwen3.5-27B-metro-v24](https://huggingface.co/continker/Qwen3.5-27B-metro-v24) | 16 GB | a 32 GB machine |
 
-[Try the live demo](https://huggingface.co/spaces/remcohendriks/metrollm) · [Browse the collection](https://huggingface.co/collections/continker/metrollm-bench-v24-6a35b586a11068e1b1ba3d47) · [Benchmark and code](https://github.com/continker/metrollm-bench)
+[Paper](https://huggingface.co/papers/2609.10016) · [Dataset (955 cases)](https://huggingface.co/datasets/continker/metrollm-bench) · [Live demo](https://huggingface.co/spaces/remcohendriks/metrollm) · [Collection](https://huggingface.co/collections/continker/metrollm-bench-v24-6a35b586a11068e1b1ba3d47) · [Benchmark and code](https://github.com/continker/metrollm-bench)
 
 ---