{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreie6hciv4edv2m7hlqfxi4wfxe3riotewdisq3bv3if7w4624dmqsa",
"uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mmuivfyt6c62"
},
"path": "/t/semantic-bundle-ai-a-complementary-layer-for-llms-91-7-memory-reduction-38-6-drift-reduction-zero-retraining/176269#post_1",
"publishedAt": "2026-05-27T21:48:58.000Z",
"site": "https://discuss.huggingface.co",
"tags": [
"Search results",
"GitHub - msaitou-glitch/Semantic-Bundle-AI: Official repository for the \"Meaning Bundle AI\" project. Complementary Layer to LLMs using Stable Coordinate Systems. · GitHub"
],
"textContent": "Hi HF community,\n\nI’d like to share two preprints on **Semantic Bundle AI** , a drop-in complementary layer for existing LLMs that addresses three structural problems: semantic drift, difficulty of targeted edits, and memory overhead.\n\n**The problem**\n\nWhen you update a concept in an LLM’s embedding space:\n\n * The change drifts across unrelated concepts (semantic drift)\n * You can’t surgically edit one concept without contaminating others\n * Storing full embeddings at scale is expensive\n\n\n\n**The approach**\n\nSemantic Bundle AI sits **on top of** existing LLMs — no architectural changes, no retraining required.\n\n * **Anchor coordinates** : stable reference frames that resist drift\n * **Semantic bundles** : structured concept representations with controlled update locality\n * **Sparse reconstruction** : compress stored embeddings via bundle-based reconstruction\n\n\n\n**PoC results (4 experiments)**\n\nMetric | Result\n---|---\nMemory reduction (K=64) | **91.7%** (45.0 KB → 3.8 KB)\nReconstruction similarity | **0.963**\nCumulative drift reduction | **38.6%**\nEdit contamination rate | **32.6%** of baseline (at ρ=0.1)\n\nZero retraining. Zero architectural modifications.\n\n**Papers & code**\n\nZenodo (Paper 0 + Paper 1): Search results\nCode: GitHub - msaitou-glitch/Semantic-Bundle-AI: Official repository for the \"Meaning Bundle AI\" project. Complementary Layer to LLMs using Stable Coordinate Systems. · GitHub\n\n**Limitations (honest)**\n\n * Small-scale controlled datasets (15–110 sentences, single domain)\n * Stability–ranking tradeoff identified (anchor coordinates improve cluster stability but not ranking consistency)\n * Not yet validated at production scale\n * Paper 1 under review at SSRN\n\n\n\nLooking for critiques, failure cases, and adjacent work. Happy to discuss.",
"title": "Semantic Bundle AI: A Complementary Layer for LLMs — 91.7% Memory Reduction, 38.6% Drift Reduction, Zero Retraining"
}