论文《Present but Hard to Read: Probing Latent Reasoning with Minimal Pairs》用最小对实验挑战了“探针读不出即模型未表征”的默认假设。