| ▲ | mycall 6 hours ago | |
What the paper contributed was choosing which parameters to change, proposing edit and testing them against graded examples. The objective was the model’s answer accuracy, which depends nonlinearly on its weights. The model guided a gradient free search over weight edits. The paper also reports a separate result in which an intact Qwen located and clipped corrupted weights in a damaged copy, partially restoring its vision performance. | ||