SCALPELLO-GDP: a frontier model edits the weights of Qwen3.5-0.8B directly. One Householder reflection of a 'repetitive re-checking' direction lifts held-out math accuracy from 9.4% to 53.6% (384 paired problems), with no training
SCALPELLO-GDP is a local replication of weight-space programming (Takehi, Teoh, Zhang, Khattab, MIT CSAIL, 2 Nov 2026). A frontier model, OpenAI gpt-6-astra called through the API, acts as an editor that reads and writes the weights of Qwen3.5-0.8B directly: no training, no gradient descent, no distillation, no retrieval; the editor itself is not modified. Each turn it writes a Python block run in an offline CPU kernel holding the weights, or requests a forward pass on a local Tesla T4 (16 GB, fp16; 25 passes; 8,192-token cap). Diagnosis: most failures were answers cut off while the model kept re-checking its work. The editor extracted a 'repetitive re-checking' direction u (residual stream, layer 20) and applied the Householder reflection W' = W − 2u(uTW) to the 12 attention output projections of layers 12–23 (9 Gated DeltaNet out_proj, 3 full-attention o_proj). Random-direction control: +3.1 ± 5.5 points. Edit frozen after pass 17. Held-out (6 disjoint cohorts of 64 practice problems, 32 GSM8K + 32 MATH train split, paired, never used for selection): accuracy 9.38% → 53.65% over 384 problems, paired gain +44.27 ± 2.56 points (95% bootstrap CI 39.32 to 49.22), 171 gains / 1 loss, exact McNemar p = 5.8e-50; answers cut at the limit 91.7% → 28.1%. Forward 18: 10.94% → 60.94% (32 gains / 0 losses) Forward 21: 14.06% → 64.06% (32 gains / 0 losses) Forward 22: 6.25% → 43.75% (24 gains / 0 losses) Forward 23: 6.25% → 43.75% (25 gains / 1 losses) Forward 24: 9.38% → 54.69% (29 gains / 0 losses) Forward 25: 9.38% → 54.69% (29 gains / 0 losses) Every forward pass: exit code 0, network audit CLEAN (0 connections). Editor API cost: $8.43. Limits: practice (train-split) problems, not MATH-500/GSM8K test; 8,192 instead of 32,768 tokens; one sample, one run. Edited model (Apache-2.0): Hugging Face odelpa/qwen3.5-0.8b-scalpello-b2-gdp (safetensors + GGUF), Ollama ollamagdp/qwen3.5-0.8b-scalpello-b2-gdp. Files: Italian, English and bilingual PDFs (with appendix listing every problem whose outcome changed), figures, scalpello_code_results_v1.0.zip (code, editor turns, requests, generation outputs, network reports, analysis; no weights inside the zip, no credentials). Author: Giovanni Del Papa, Independent researcher, Italy. IT — SCALPELLO-GDP: gpt-6-astra (via API, non modificato) modifica direttamente i pesi di Qwen3.5-0.8B con una riflessione di Householder della direzione del «ricontrollo ripetitivo» in 12 proiezioni di uscita dell'attenzione (strati 12–23), senza addestramento. Su 384 problemi held-out appaiati: 9,38% → 53,65% (+44,27 ± 2,56 punti; 171 migliorati, 1 peggiorato); risposte interrotte 91,7% → 28,1%. Rete CLEAN, costo API 8,43 $. Modello pubblico: HF odelpa/qwen3.5-0.8b-scalpello-b2-gdp, Ollama ollamagdp/qwen3.5-0.8b-scalpello-b2-gdp. Version 1.1 — 21 Nov 2026: changes to the PDFs only; data, results, numbers and figures are unchanged. (1) Appendix A now gives the full text of every problem whose outcome changed (v1.0 truncated it to 150 characters), taken from the run's practice files and checked to be identical to the original GSM8K train and MATH train items; the table was redesigned (landscape pages, separate columns for the reference answer, whether the edited model is correct, the edited model's answer and the original model's answer, with answers cut at the 8,192-token limit marked explicitly). (2) New plain-language section 'Why the model improved' / 'Perché il modello è migliorato', with two new bibliography entries (Qwen3.5-0.8B model card; Li et al. 2026, arXiv:2609.37326); existing bibliography entries are unchanged. (3) problemi_migliorati.csv/.md and problems.py inside the code zip now contain the full problem texts (zip file name unchanged).
Authors
- Giovanni Del Papa
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24416417
- Primary Topic
- Topic Modeling
- Type
- preprint