从平均统计那一行可以看到两个backbone加上On-Adapter之后$p<0.001$满足$p<...

可以把逻辑进一步收紧为“总体显著性 → 不同插入位置的差异 → 对机制的解释”，这样更像论文中的实验分析。

润色后的中文：

:::writing{variant="document" id="18427" title="实验结果分析（中文润色版）"} 从平均统计结果可以看出，在两个 backbone 上引入 On-Adapter 后均取得了显著提升（$p<0.001<0.05$），表明 On-Adapter 能够稳定提升模型性能。进一步地，不同 backbone 在不同插入位置上的收益存在明显差异，这可能与各层隐变量的表征能力有关。对于 GPT4MTS，在浅层之前接入具有在线预测能力的 On-Adapter 能够带来显著提升，而在浅层之后接入时，其增益不再显著。相比之下，在 VoT 中，On-Adapter 在浅层和中间层的插入均能获得显著收益，而在网络头部和尾部的提升则不显著。上述结果表明，On-Adapter 的效果与中间隐变量的表征质量密切相关，同时也进一步验证了其在线预测机制的有效性。 :::

顶会风格英文版：

:::writing{variant="document" id="73164" title="Top-Conference English Version"} As shown by the averaged results, incorporating On-Adapter leads to statistically significant improvements on both backbones ($p<0.001<0.05$), demonstrating its consistent effectiveness. Interestingly, the performance gains vary substantially across insertion positions, suggesting that the effectiveness of On-Adapter is closely related to the representational quality of latent features at different depths. For GPT4MTS, introducing On-Adapter before the shallow layers yields significant improvements, whereas inserting it after the shallow layers does not result in statistically significant gains. In contrast, for VoT, significant improvements are observed when On-Adapter is inserted at the shallow and intermediate layers, while the gains at the input-side and output-side positions are not statistically significant. These observations highlight the importance of latent representation quality and further validate the effectiveness of the online prediction capability introduced by On-Adapter. :::

如果用于 NeurIPS / ICML / ICLR 的 ablation 或 significance analysis，还可以再写得更“机制导向”，减少重复出现 significant improvements，让论述更像论文正文。