Skip to content

THOX.ai Research · Paper v1.1.0 · 2026-11-04

Preprint, not peer reviewed

Weight-Space Programming for THOX Edge Models: Evidence Review, Small-Scale Recreation, and a Signed Patch Channel

Craig Ross

Chief Executive Officer

ORCID 0009-0005-9872-3673

Tommy Xaypanya

CTO and Co-Founder (corresponding author)

ORCID 0009-0000-4425-9565

THOX.ai LLC · Cedar Park, Texas · November 4, 2026

DOI 10.5281/zenodo.23955394

Abstract

Takehi et al. report that a frontier LLM can read and write neural network weights directly, without gradient descent, and that editing 0.016% of Qwen3.5-0.8B’s weights raised its MATH-500 score from 7.9% to 48.1% by removing a thinking-loop failure. THOX ships a derivative of that exact base model, so this report reviews the claim, recreates its three verbs at checkable scale, re-verifies the released edit and specifies the control the question turns on. The headline edit is genuine and narrower than its framing: it removes an obstacle rather than adding ability. The decisive control, the same base model with thinking turned off, is specified at full MATH-500 scale with a pre-registered decision rule; its run was stopped by compute billing before completion, so this edition states no number for it. The released edit rebuilds bit-exactly (130,188 entries, both hashes), but a factor form estimated from it is not exact without a sparse residual, which sets the size of an exact patch at tens of kilobytes. Small recreations show task identity is legible only in composed per-class circuits, that a hand-written network matches an oracle-trained one, and that 16 of 747 weights repair a stuck policy. THOX adopts the decoding defences and the signed patch channel now, and holds weight patching behind a pre-registered bar: beat the best decoding-only control by more than two standard errors.

130,188

Entries changed in the released Qwen3.5-0.8B edit, rebuilt bit-exactly and re-verified on a second host (E0-R)

Measured

100%

Task identified from weights alone by the per-class composed reader, vs 33.3% (chance) for three marginal readers (E1, 30 networks)

Measured

306 / 0

MATH-500 problems where the released edit scores higher / lower than the base, in the authors’ own per-problem files (A0)

Calculated

80.6 KB

Exact patch for the released edit (factor plus sparse residual), vs 387.6 KB for the authors’ released sparse file (P1, corrected in v1.1.0)

Measured

Download the paper (PDF)Read the paper (HTML) Browser demo

Contents

  1. 1. Introduction
  2. 2. The source result and its evidence state
  3. 3. Definitions and method
  4. 4. THOX recreation experiments
  5. 5. The decoding control: protocol and decision rule
  6. 6. The THOX patch channel
  7. 7. Application to THOX device models
  8. 8. The gated replication program
  9. 9. Limitations
  10. 10. Conclusion
  11. 11. Provenance and document control
  12. References

Where the paper lives

Cite this paper

@misc{ross_craig_2026_wsp,
  author       = {Ross, Craig and
                  Xaypanya, Tommy},
  title        = {{Weight-Space Programming for THOX Edge Models:
                   Evidence Review, Small-Scale Recreation, and a
                   Signed Patch Channel}},
  month        = nov,
  year         = {2026},
  publisher    = {THOX.ai Research},
  version      = {v1.1.0},
  doi          = {10.5281/zenodo.23955394},
  url          = {https://thox.ai/research/papers/weight-space-programming}
}

Preprint, not peer reviewed; DOI 10.5281/zenodo.23954379.

Your AI. Your Data. Your Rules.™