Spaces:

---

vigneshwar234

/

TemporalMesh-Transformer-Demo

like 1

Running

Try the TMT Live Demo

This Space lets you run a full TMT forward pass in your browser — no GPU needed.

What you can do:

• Type any text → watch the $k$NN attention graph evolve across all 12 layers
• See exactly which tokens exit early (punctuation → layer 2) vs late (rare words → layer 12)
• Explore the ablation benchmark charts (PPL vs compute for all 8 configurations)
• Read the full paper and copy the citation

Demo tabs:

1. ⚡ Live Forward Pass — dynamic graph topology + exit gate confidence heatmap
2. 📊 Benchmarks — ablation bar chart, Pareto curve, complexity scaling, exit depth by token type
3. 🏗 Architecture — full technical description of all 5 innovations
4. 📄 Paper & Citation — BibTeX, DOI, Zenodo link

The forward pass is implemented in pure NumPy (no PyTorch dependency) so it runs instantly on CPU.

Paper: https://zenodo.org/records/20287390
Code: https://github.com/vignesh2027/TemporalMesh-Transformer