Locality-Preserving Multiscale Computation for Compact Language Models

Transformer inference repeatedly moves and compares large amounts of state. Sparse attention, retrieval, caching, quantization, and small models reduce this burden, but their gains are often designed independently. The Glyphd fractal-computation program asks whether they can be coordinated through a shared multiscale geometry.

The proposal combines space-filling curves, recursive tilings, dynamic spatial indexes, foveated context selection, semantic object graphs, and a Q-Frontier that identifies the smallest unresolved set requiring expensive reasoning. Local detail is processed densely; increasingly distant or settled regions are represented more coarsely. Hilbert, Z-order, or Gosper-style orderings provide deterministic one-dimensional traversal of multidimensional neighborhoods. The geometry does not compute answers and does not compress arbitrary data. It organizes locality, candidate sparsity, cache layout, and evidence-preserving omission.

The program has two distinct levels. The first is practical runtime architecture around a compact model: structured outputs, PromptCapsules, prefix reuse, foveated retrieval, small deltas, and software scheduling. The second is model-architecture research: multiscale sparse attention and locality-preserving state layout. The first can be benchmarked now. The second requires kernels, trained models, and matched baselines before any performance claim.

Current evidence state

SPECIFIED

This label describes the strongest public-safe state of the specific paper claims. Open the paper's claim ledger for claim-level distinctions and limitations.

Publication

• Full paper: /papers/p13-fractal-llm-computation
• Claim ledger: content/papers/p13-fractal-llm-computation/claim-ledger.json
• Source manifest: content/papers/p13-fractal-llm-computation/source-manifest.json
• Diagrams: 4
• Approximate body length: 1,803 words

Demonstrations

• d09-fractal-llm-locality

Benchmarks

• b05-surfacedelta-uplift
• b06-jeffe-4b

Public boundary

The research entry publishes a public-safe architecture and evidence record. It does not grant implementation rights or expose restricted claim-ready material.