The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment

Benchmark Model Rank Results
mathematical-reasoning-on-minerva-mathQwen3 (14B) + UNLOCK from 4B#25Accuracy: 39.2
mathematical-reasoning-on-olympiadbenchQwen3 (14B) + UNLOCK from 4B#45Accuracy: 36.3
math-word-problem-solving-on-mathQwen1.5 (14B) + UNLOCK from 7B#133Accuracy: 31.2
arithmetic-reasoning-on-gsm8kQwen1.5-14B + Unlock#177Accuracy: 74.4