Published March 24, 2026
| Version v2.0.0
Software
Open
augchan42/king-wen-agi-framework: v2.0.0: Negative Result Paper with Experimental Data
Authors/Creators
- 1. IBM Research
- 2. @synestize, @csiro-mlai
- 3. Wikimedia Foundation
- 4. Politecnico di Milano
- 5. University of Maryland, College Park
- 6. @qassay
- 7. @TheochemUI, @lab-cosmo, @metatensor
- 8. Los Angeles Dodgers
- 9. Chalmers University of Technology and University of Gothenburg
Description
Major revision: Paper rewritten as negative-result report
The King Wen sequence has genuine statistical properties (confirmed by Monte Carlo analysis) but they do not improve neural network training.
What changed from v1.x
- New experimental results from autoresearch (PyTorch/CUDA) and autoresearch-mlx (Apple Silicon)
- Finding: anti-habituation properties don't help ML training. LR modulation degrades performance. Curriculum ordering is worst non-sequential on CUDA, within noise on MLX.
- Fixed corrupted King Wen binary encoding (10 duplicate hexagrams corrected)
- Regenerated all statistical results with validated encoding
- New paper title: "Statistical Properties of the King Wen Sequence: An Anti-Habituation Structure That Does Not Improve Neural Network Training"
- Removed speculative sections, OpenSpiel proposal, and Claude endorsement
- Added Limitations section acknowledging scale constraints
- Full LaTeX + PDF ready for arXiv submission
Paper (PDF)
Built with XeLaTeX from paper/king-wen.tex. 9 pages including appendices.
Files
augchan42/king-wen-agi-framework-v2.0.0.zip
Files
(1.1 MB)
| Name | Size | Download all |
|---|---|---|
|
md5:0a6e32637605cb5561573dea0a5f7626
|
1.1 MB | Preview Download |
Additional details
Related works
- Is supplement to
- Software: https://github.com/augchan42/king-wen-agi-framework/tree/v2.0.0 (URL)
Software
- Repository URL
- https://github.com/augchan42/king-wen-agi-framework