Feedback Loop Collapse

Model relies on irrelevant signals.

Small prompt changes cause large output changes.

Probabilities do not reflect true correctness.

Applying learned patterns incorrectly.

Differences between training and inference conditions.

Loss of old knowledge when learning new tasks.

Train/test environment mismatch.