Rohan Paul (@rohanpaul_ai)T2·78 pts
- The model edits its own weights via reasoning with gradients blocked, bypassing training clusters and extra inference compute
- Accuracy rose from 57.2% to 63.8% on 490 pre-registered held-out problems
- There is a clear scale threshold: 2B produced no edit, while 4B and 9B degraded themselves
- Practical guidance is to keep only self-edits that pass held-out tests