Strands Decider：为何未采用编码器架构

strands-decider-2B模型的架构设计选择成为研究焦点。1通过对比五种设计方案，包括纯编码器e1b、编码-解码器混合等架构，分析显示编码器在延迟性能上具有明显优势——e1b的中位延迟约35毫秒，相比采用GatedDeltaNet架构的v19降低46%（v19延迟为65毫秒）。1

然而，性能对比结果更为复杂。1e1a在九项任务中的七项上超越v19，但在JevBench任务上表现显著较差（153分对比v19的168分）。1从架构角度看，v19的GatedDeltaNet架构占其层数的75%，具有线性计算成本，而e1b则包含26层全注意力层，具有二次方计算成本。1这一差异导致e1b在处理较长提示长度时扩展性能更差，与其注意力机制的二次方成本表现一致。1综合考量延迟、准确率和泛化能力等因素，编码器和解码器架构在该任务规模下均未呈现出明显的绝对优势。1

The Strands Decider team's choice to build strands-decider-2B around an improved decoder architecture rather than a pure encoder design reflects a nuanced trade-off between competing priorities, according to an architectural analysis.1 While encoder designs like e1b demonstrate significant latency advantages—achieving a median latency of approximately 35 milliseconds compared to 65 milliseconds for the decoder-based v19 variant, a 46 percent reduction1—the performance comparison across accuracy and generalization reveals no clear winner between the two approaches.1

The architectural difference stems from fundamental computational trade-offs. The v19 decoder employs a GatedDeltaNet architecture that comprises 75 percent of its layers and operates with linear computational cost, while the e1b encoder incorporates 26 fully attentional layers that incur quadratic computational costs.1 In testing across nine tasks, e1b surpassed v19 in seven of them, but showed notably worse performance on JevBench, scoring 153 compared to v19's 168.1 The encoder's scaling limitations become apparent with longer prompt lengths, a weakness consistent with the quadratic cost structure of its attention mechanism.1 For tasks at this model scale, neither encoder nor decoder architecture emerges as an unambiguous winner, suggesting that the decoder choice represents a deliberate optimization rather than a technical inevitability.1