buk81/uniformity-asymmetry: Release v2.0 - Phase-Structured Dynamics

Paper #2 Published Layer-wise Embedding-Output Dynamics Across LLM Families: Evidence for Phase-Structured Decision Commitment Key Findings | Model | Type | Early Layers | Late Layers | Phase-Structured? | |-------|------|--------------|-------------|-------------------| | Pythia-6.9B | Base | +0.44*** | **-0.17*** | Yes | | Llama-3.1-8B | Base | +0.05 | **-0.30*** | Yes | | Apertus-8B | Multilingual | +0.39*** | **-0.25*** | Yes | | Gemma-2B | SFT | +0.10 | -0.02 | No (boundary) | Central Conclusion "Decision commitment in large autoregressive transformers is not uniformly localized at the final layer, but emerges as a late-stage inversion whose depth and sharpness depend on scale, supervision, and architecture." What's New in v2.0 📄 Paper #2: Full LaTeX source + PDF in paper2/ 📓 7 Notebooks: Colab-ready validation for all 4 model families 📊 Results: JSON data for Pythia, Llama, Gemma, Apertus ⏱️ Timestamps: Bitcoin blockchain proofs for both papers Papers | # | Title | DOI | |---|-------|-----| | 1 | Uniformity Asymmetry | 10.5281/zenodo.18110161 | | 2 | Phase-Structured Dynamics | 10.5281/zenodo.18142454 | Relation: Paper #2 continues Paper #1, addressing output correlation and layer-wise analysis. Citation @article{delia2026phase, author = {D'Elia, Davide}, title = {Layer-wise Embedding-Output Dynamics Across {LLM} Families: Evidence for Phase-Structured Decision Commitment}, journal = {Zenodo preprint}, year = {2026}, doi = {10.5281/zenodo.18142454} }