{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreignbihmjcrcpkcenodlkym7ae5o2cosjblxel63qfnmknqslhggfu",
"uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mmku2ykvqqm2"
},
"path": "/t/continuation-drm-transformer-from-open-geometry-to-negotiated-geometry-in-ai-alignment/176106#post_5",
"publishedAt": "2026-05-24T00:13:13.000Z",
"site": "https://discuss.huggingface.co",
"textContent": "Thanks, Jean. I agree that the right next step is a matched perturbation + recovery test.\n\nOne clarification: DRM has already reached a T²-like toroidal closure regime at ~3.5M parameters. It is not yet strict stable closure under the current criterion, since stable remains <= 0.60, but the toroidal signature is already there.\n\nSo the contribution is not that DRM has fully stabilized closed geometry yet. The contribution is that a very small model, trained from the beginning inside a geodesic/relational manifold, can already enter an unstable or near-stable T²-like regime.\n\nThat is exactly why I think the architecture matters. Standard Transformers may exhibit measurable emergent geometry, but DRM is testing whether topology can be induced directly and then stabilized.",
"title": "[Continuation] DRM Transformer: From Open Geometry to Negotiated Geometry in AI Alignment"
}