10.5.5
Diagnosing Capability Regression
Use base-model comparisons, prompt tests, perplexity checks and few-shot recovery to identify the source of performance degradation.
What is it?
Capability regression can be diagnosed using base-model comparisons, prompt engineering, perplexity and few-shot recovery. These tests help distinguish erased knowledge from knowledge that remains but is difficult to trigger under the aligned policy.
Key points
Base-model comparison: test the failing tasks on the original checkpoint and fine-tuned model.
Prompt engineering: recovery through careful prompting suggests alignment tax.
Perplexity: a spike on the pre-training validation set suggests catastrophic forgetting; stable perplexity suggests alignment tax.
Few-shot recovery: recovery after in-context examples suggests alignment tax.
Failure to recover even with many examples suggests catastrophic forgetting.
A Practical Capability Regression Diagnosis
1. Run the same failing task on the base model and fine-tuned model.
2. Try a carefully designed prompt to see whether the fine-tuned model can recover the capability.
3. Compare perplexity on the pre-training validation set.
4. Provide a few in-context examples and test again.
Interpretation:
- Base model succeeds but fine-tuned model fails completely: investigate catastrophic forgetting.
- Prompting or few-shot examples restore performance: alignment tax is more likely.
- Perplexity spike: evidence of forgetting.
- Stable perplexity: evidence consistent with alignment tax.
Real-world application
Build a regression test suite that compares base and fine-tuned checkpoints. Include representative prompts, pre-training validation perplexity and few-shot recovery tests before deciding whether to adjust SFT or alignment training.
Key takeaway
Diagnose whether knowledge was erased or suppressed before attempting to fix capability regression.