After distillation your 8B model recovers MMLU well but...

Multi-step reasoning often depends on specific layers within the residual stream that accumulate state. If layers were pruned based on general cosine similarity, reasoning-critical layers may have been inadvertently removed. To fix this, re-run importance scoring using a calibration set that includes chain-of-thought traces to identify reasoning-specific layers. Restore these layers, remove less-critical ones to maintain the parameter budget, and perform a new round of distillation using a reasoning-focused corpus.