WeightSpaceProgramming

Home replication attempt of MIT CSAIL's "Weight-Space Programming" (Nov 2, 2026): a frontier LLM edits Qwen3.5-0.8B's weights to stop its math reasoning loops, with no gradients and no forward passes. Here the editor is Claude instead of GPT-6 Astra.

Setup

• Model (original, read-only): ~\tools\models\Qwen3.5-0.8B sha256 04b1c301231dd422b8860db31311ab2721511346a32cb1e079c4c4e5f1fe4696
• Python env: uv sync (uv at ~\tools\uv\0.12.19\uv.exe). Torch here is CPU-only and is used for weight math only.
• Inference/eval: llama.cpp at ~\tools\llama.cpp\b11804 (GTX 1070), via GGUF conversions.
• Paths: scripts find models, llama.cpp and its source under ~/tools; set WSP_TOOLS to use another folder. pyproject.toml expects llama.cpp's gguf-py at ../../tools/source-code-of-open-source/llama.cpp-b11804/gguf-py relative to the repo.

Protocol

1. Baseline: MATH-500 subset, record accuracy and loop rate (hit token limit). Run both greedy and Qwen's recommended sampling (presence_penalty 1.5).
2. Blind edit: analyze weights with numpy/torch, no forward passes. Every analysis step logged in logs/. Edits are written only to copies in edited/.
3. Re-test edited model plus a same-size random-direction control.

Layout

src/ scripts, runs/ eval outputs, one folder per run at runs/<model>/<task>/<run>/ (results.jsonl, summary.json with the exact settings, console.log, per-problem transcripts in transcripts/, rendered Markdown regenerated from results.jsonl by src/export_transcripts.py; long ones are split into _p2, _p3 parts because GitHub renders at most about 1950 math expressions per page), with runs/INDEX.md as the table of every run, edited/ modified weight copies (gitignored), logs/ edit-session log, docs/ notes.

Current state, plan and results: docs/STATUS.md.