Failure Modes: Capacity Mismatch, Bias Amplification, and Distribution Drift

Knowledge distillation fails in predictable ways when assumptions break. Capacity mismatch occurs when the student is too small for task complexity, causing underfitting even with perfect teacher signals. Symptoms include high bias on long tail inputs, poor calibration with high expected calibration error (ECE) above 0.1, and large gaps between teacher and student on rare classes. For example, distilling a 12 layer natural language understanding model to 2 layers might maintain 90 percent average accuracy but drop to 60 percent on rare intents that require deeper reasoning. The remedy is increasing student width or depth, or switching to task specific feature distillation that preserves critical representations.

Teacher errors and bias amplification are insidious. The student inherits all teacher mistakes and biases, which can be amplified when distilling over large unlabeled corpora where systematic bias exists. If a teacher model is miscalibrated and overconfident, producing 0.95 probability for ambiguous examples, the student will be even more overconfident. With pseudo labeling on 100 million unlabeled examples, bias on protected attributes or minority groups can scale linearly with corpus size. Mitigation requires a holdout of human labeled examples for validation, explicit debiasing objectives, and post training calibration using temperature scaling or isotonic regression. Monitor not just aggregate accuracy but also fairness metrics like demographic parity and equalized odds across subgroups.

Distribution shift and model staleness create ongoing operational challenges. Teachers trained on historical data may generalize poorly to current traffic. Students with tight capacity budgets drift faster when input distributions shift, since they have less headroom to adapt. A ranking model distilled on pre pandemic user behavior might degrade 15 to 20 percent when behavior shifts abruptly. Sequence tasks like machine translation face exposure bias: token level distillation from teacher outputs propagates errors during student decoding. Multi label and metric learning tasks break with naive softmax KL, since the loss assumes mutually exclusive classes. For retrieval, you need sigmoid based losses and relation based distillation that preserve pairwise distances, or students lose ranking quality while matching marginal distributions. Privacy leakage is also real: if the teacher memorizes sensitive training strings, naive distillation can replicate memorization. Add differential privacy noise, filter high confidence outputs, or reject memorized sequences during distillation to prevent leakage.

💡 Key Takeaways

•Capacity mismatch shows as high expected calibration error above 0.1 and 30+ percent gaps on rare or complex inputs when student is too small, needs wider or deeper architecture

•Bias amplification scales with unlabeled corpus size, requiring holdout validation, debiasing objectives, and post training calibration with temperature scaling to maintain fairness metrics

•Distribution drift hits tight capacity students harder, causing 15 to 20 percent degradation on shifted traffic, needs regular redistillation cadence and replay of recent production data

•Sequence tasks suffer exposure bias from token level distillation, sequence level distillation with sampled or beam outputs helps but can overfit to teacher decoding strategy

•Multi label and retrieval tasks need sigmoid or ranking losses, not softmax KL, plus relation based distillation to preserve distances or students lose retrieval quality by 10+ percent Mean Average Precision

📌 Examples

Ranking model capacity failure: distilling 12 layer teacher to 2 layer student maintains 88 percent NDCG on head queries but drops to 65 percent on tail, expected calibration error rises from 0.05 to 0.18, fixed by using 4 layer student

Bias amplification case: teacher with 2 percent gender bias on resume screening, distilled on 10 million unlabeled resumes, student shows 8 percent bias, mitigated by adding 100,000 human audited examples and demographic parity constraint

Text embedding drift: student distilled on 2022 data, serving in 2024, loses 12 percent Mean Average Precision on current queries, fixed by quarterly redistillation with recent 6 month traffic window

← Back to Knowledge Distillation Overview