Comparing Transformers and Hybrid Models at the Token Level
摘要
作者利用 Olmo 3 和 Olmo Hybrid 的开放权重,在相同前缀和目标 token 上对比纯 Transformer 与混合模型(注意力 + 循环层)的损失。混合模型在多数词性标签上损失更低,但增益不均:对开放类实词增益最大,对封闭类功能词增益较小;在散文、代码和标记语言中,对开分隔符的损失优势大于闭分隔符,且在重复 n-gram 上优势几乎消失。合成探针任务显示,混合模型在代词记忆和实体追踪任务上占优,而 Transformer 在括号匹配(选择闭分隔符)任务上占优。这表明循环层有助于利用文档语义状态进行预测,而注意力更擅长基于 n-gram 复制或句法括号匹配的预测。作者还展示了基于 token 级分解的筛选评估如何锐化混合架构的预训练诊断。
荐读理由
你可以据此判断:混合架构的优势集中在需要追踪文档语义状态的任务(如代词指代、实体跟踪),而在依赖局部模式复制或句法匹配的任务上,纯Transformer可能更合适。
原文(据论文摘要)
Hybrid language models that mix attention and recurrent layers have shown promise: theoretically, recurrent layers ameliorate the limitations of pure transformers on state tracking, and empirically, hybrids can outperform pure transformers in loss and downstream evaluations \citep{waleffe2024empirical,merrill2026olmohybrid}. Yet it remains unclear which data or capabilities drive these gains, and to what degree they reflect the theoretical advantages motivating hybrid models. We address this question using the open weights from Olmo 3 \citep{olmo2025olmo3} and Olmo Hybrid \citep{merrill2026olmohybrid}: we compare the loss of a matched transformer and hybrid at the same target tokens under the same prefixes, stratifying the results by natural token tags, copy features, delimiter structure, and controlled synthetic probes. The hybrid has lower loss on most tag families, but the gains are not uniform: they are largest for open-class content words and smaller for many closed-class function words. Across prose, code, and markup, the hybrid's loss advantage is larger on opening delimiters than on the corresponding closing delimiters, and nearly vanishes on repeated $n$-grams. Synthetic probes show the same split: the hybrid is favored on pronoun-memory and entity-tracking tasks, whereas the transformer is favored on bracket-matching tasks that require choosing closing delimiters. These patterns suggest that the recurrent layers in hybrids improve predictions that leverage the semantic state of a document, whereas attention helps on tokens predictable by $n$-gram copying or syntactic bracket matching. We conclude with proof-of-concept filtered evaluations showing how token-level decompositions can sharpen pretraining diagnostics for hybrid architectures.
这条对你有帮助吗?