Awesome Weight-Space Programming
Give a language model another network's weights and let it change the numbers directly to explain, repair or build behaviour, with no gradient descent. I track this new field daily: the papers, code, benchmarks and critiques, plus eighty years of earlier work that edited weights under other names.
- Curated research map · 2026
- Updated daily
- Python build + JSON data
- CC0
What is weight-space programming?
The term comes from the MIT CSAIL report Sparks of Weight-Space Programming (2026-11-02). A general LLM is handed a network's weights, sometimes with code and a budget of forward passes, and edits the numbers itself. In that report a model names what networks were trained on from their weights alone, repairs a stuck control policy by changing a few percent of its weights in one reply, and writes small game-playing networks from zero.
- ReadPredict what a network does, or what it was trained on, from its weights.
- RepairFix a specific behaviour by editing weights directly.
- WriteConstruct a working network from scratch, number by number.
How fast it is moving
New entries since the term was coined
Everything in the list, by year
Explore the list
Search and filter
Loading…This explorer reads data/entries.json straight from the GitHub repository, so it shows the same entries as the list itself. If GitHub is unreachable it falls back to a snapshot saved with this page.