7.9% → 48.1%MATH-500 accuracy
Editing a small fraction of Qwen3.5-0.8B weights reduced reasoning loops in the original research experiment.
AXONRY / WEIGHT-SPACE PROGRAMMING
Read the weights. Rewrite the behavior.
Explore direct model editing on Bittensor.
01 / READ
Give a model an objective. Let it identify the structures in a policy network that shape behavior.
02 / WRITE
Turn an objective into a local weight update. A small parameter change can shape a new policy.
03 / EXECUTE
Apply the patch and run the obstacle task again. Watch a local weight update become a new behavior.
04 / VERIFY
Axonry rebuilds the edited model and measures it against the current champion.
AXONRY / BITTENSOR SN99
Our vision is decentralized LLM training. Today, miners propose ways to improve a shared language model through direct weight edits, without gradient updates.
Arbos runs agents in parallel. Each agent tests a miner’s method in a dedicated GPU sandbox and returns a weight edit.
Axonry brings Bittensor miners together under shared rules, schedules their experiments, and benchmarks each edit against the current champion.
The long-term goal is thousands of isolated experiments running in parallel.
7.9% → 48.1%MATH-500 accuracy
Editing a small fraction of Qwen3.5-0.8B weights reduced reasoning loops in the original research experiment.
4 / 4Game policies identified
Astra inferred the objectives of four policies from their weights and input and action definitions.
3 / 3Games above the PPO baseline
Three policies written directly by Astra outperformed the trained baseline in MinAtar.