Unedited
Qwen3.5-0.8B as released
Write a problem and run both models.
hedwigaihedwigaihedwigai
Does the weight edit stop the loop?
Continue in
An AI agent edited 131,072 weights of Qwen3.5-0.8B, and its MATH-500 score with thinking on rose from 7.9% to 48.1%. Pick a problem, or write your own. Both models answer it with the same seed, and under each answer is the edit’s signal, which climbs when the model starts to repeat itself. The post explains how it works.
1. Pick a problem
2. The edit’s signal
The signal appears here once both models have answered.
How much the edit takes out of layer 19 at each token, as the unedited model reads each answer. Shaded stretches are mostly repeats, and the dashed line marks where a loop began. The edited model removes exactly this signal.
3. The problem
Loading the recorded runs…
Sign in to run
These four problems are free to look at; sign in to run your own.
4. Compare
Edited
Two layers, one direction taken out
Write a problem and run both models.