← All questions
AgentsHard· In-Context Learning· Debugging· Classification

Your four-way classifier predicts label A on 70% of a b...

A strong diagnostic approach starts with a content-free probe: feed N/A in place of the input and observe the label distribution. If the model outputs label A with high frequency even without content, it is reporting a prior bias. The fix is a simple division or diagonal correction to the output, which is significantly more efficient than resampling demonstrations or fine-tuning the model.

Code playground (Python)