feature-alignment

该论文提出 PORTAL，一种无需配对图像-文本数据、基于最优传输的多模态投影器初始化方法，可以完全跳过投影器训练阶段，在六个 LLM 主干和 16 个基准上匹敌使用 558k 图文对的标准训练流程。

This paper proposes ST-Merge, a steerable model merging framework that uses a gated cross-attention mechanism to adaptively modulate contributions of a multilingual model and a reasoning model, outperforming fixed merging approaches on multilingual reasoning benchmarks across 21 languages.