Claude Opus 5.5 edits Tesseract Hebrew weights

discuss.huggingface.co1 min read

Researchers applied weight-space programming using an AI agent to directly edit the weights of Tesseract’s Hebrew text recognizer without gradients or a training loop. The agent reduced modern Hebrew text errors from 4.18% to 2.64% on average, outperforming a last-layer refit but falling short of a full fine-tune. The authors note that the agent could explain every change it made, identifying start-up faults in the bidirectional LSTM. Open article →