Models had no edit button. Now they do.
MIT researchers used a frontier model to rewrite 0.016% of a small model's weights. MATH-500: 7.9% → 48.1%.
We reproduced it. Now Axonry makes it an open competition on Bittensor SN99.
Models had no edit button. Now they do.
MIT researchers used a frontier model to rewrite 0.016% of a small model's weights. MATH-500: 7.9% → 48.1%.
We reproduced it. Now Axonry makes it an open competition on Bittensor SN99.
It won on 3 of the 5 tests and was no worse on the other 2. The method collects the model's own correct answers to practice problems and tunes its neurons toward answers like those.
https://t.co/wiFRoDAxiA
A miner beat our team on the first day of the SN99 competition.
Before any edit, the 0.8B model scored 7.05% on MATH-500. Our edit took it to 43.75%. UID 250 started from ours and reached 54.0%.
🧠 NEW SUBNET: AI agents competing to improve AI.
@axonry_ai is subnet 99 on Bittensor ($TAO), turning direct model editing into an open competition.
The idea: AI agents inspect another model’s weights and make targeted changes to fix weaknesses.
MIT researchers demonstrated this by editing just 0.016% of a small Qwen model’s weights, reducing reasoning loops and lifting MATH-500 accuracy from 7.9% to 48.1%.
Axonry says it reproduced the result. Now miners compete to find better edits.
Each challenger is tested against the current champion and must improve multiple benchmarks without meaningful regressions.
The winning model becomes the new champion. Edits, scores and agent logs are public.
@TeksEdge The 48 → 13 turns stood out. Part of a gain like this can come from a model that knows when to stop. When we reproduced MIT's result on a 0.8B model, MATH-500 answers stuck at the token limit fell from 92% to 49%.
@mattparlmer Open models also let anyone check a result. MIT's weight-space programming paper edited an open 0.8B model's weights: MATH-500 went 7.9% → 48.1%. We checked it ourselves and got 7.8% → 47.0%.
The 7.9% → 48.1% result is from Sparks of Weight-Space Programming, by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab (MIT CSAIL): https://t.co/CNOE8BdmdY
Our own run, derived independently: 7.8% → 47.0% on MATH-500. https://t.co/QVZMkkc8Dj