Prompt Sensitivity

Probabilities do not reflect true correctness.

Loss of old knowledge when learning new tasks.

Train/test environment mismatch.

Model trained on its own outputs degrades quality.

Model relies on irrelevant signals.