Knowledge Distillation Asymmetrically Affects Small LLM Bias

This research reveals that applying knowledge distillation to small instruction-tuned language models has uneven effects on their biases. Specifically, while distillation from a larger teacher model (Gemma-2-9B) can improve context-following on unambiguous tasks, reducing error rates, it simultaneously harms the model's ability to correctly refuse ambiguous questions on a per-item basis. For instance, items where a baseline model correctly abstained instead received stereotypical answers after distillation, even if the overall refusal rate remained similar. This pattern was consistent across different student model families. The study attributes this calibration loss to a data-side mechanism, noting that training corpora contain very few examples of refusal-as-answer-shape. Attempts to inject refusal examples during supervised fine-tuning either broke parsing or led to an over-correction, resulting in models that trivially refused almost all queries. The findings underscore the inadequacy of aggregate stereotype metrics, which can obscure this asymmetric harm. The paper proposes a new protocol, Per-Condition Calibration Diagnosis (PCCD), to accurately evaluate refusal calibration, context-following, and capability preservation, catching nuanced failures that traditional metrics miss.