Executing Writer
A small writer model used to reverse-recover skill documentation from LoRA adapters; trained on an adapted model, it reconstructs readable textual instructions by executing tasks.
Timeline (last 90 days)
The Executing Writer is a small writer model built on top of the adapted model, and on held-out skills it recovers 39-44% of key instructions.
As a control, when using the adapter of another skill, the Executing Writer recovers only 1.3-1.5% of key instructions.
On an adapter adapted to 41 external skills, the documents recovered by the Executing Writer can restore 85-91% of the original task gains.
For adapters trained only on behavior (without storing document content), the trained writer recovers only 1-2% of instructions.
The most effective path to recovering skill documents from a LoRA adapter and a frozen base model is to have the model first perform the task and then reconstruct the text from behavioral differences, rather than decoding directly from the weights.
All Facts (5)
The Executing Writer is a small writer model built on top of the adapted model, and on held-out skills it recovers 39-44% of key instructions.
50%UnverifiedAs a control, when using the adapter of another skill, the Executing Writer recovers only 1.3-1.5% of key instructions.
50%UnverifiedOn an adapter adapted to 41 external skills, the documents recovered by the Executing Writer can restore 85-91% of the original task gains.
50%UnverifiedFor adapters trained only on behavior (without storing document content), the trained writer recovers only 1-2% of instructions.
50%UnverifiedThe most effective path to recovering skill documents from a LoRA adapter and a frozen base model is to have the model first perform the task and then reconstruct the text from behavioral differences, rather than decoding directly from the weights.
50%