📄 中文摘要
递归推理模型如层次推理模型（HRM）和微型递归模型（TRM）表明，小型共享权重网络能够通过迭代精炼潜在状态来解决计算密集型和NP难题。然而，它们的训练通常依赖于深度监督和/或长时间展开，这会增加实际计算时间并可能导致模型偏向贪婪的中间行为。递归干茎模型（RSM）提出了一种递归推理方法，保留了TRM风格的骨干，同时改变了训练契约，使网络学习到一个稳定的、与深度无关的转移算子。RSM在训练过程中完全脱离隐藏状态历史，将早期迭代视为独立的“热身”步骤，并仅在最后一步应用损失函数。
形式遵循功能：递归干茎模型
出处: Form Follows Function: Recursive Stem Model
发布: 2026年3月18日
📄 English Summary
Form Follows Function: Recursive Stem Model
The Recursive Stem Model (RSM) is introduced as a recursive reasoning approach that retains the backbone of the Tiny Recursive Model (TRM) while modifying the training contract to enable the network to learn a stable, depth-agnostic transition operator. Unlike traditional recursive reasoning models such as the Hierarchical Reasoning Model (HRM) and TRM, which rely on deep supervision and long unrolls, RSM fully detaches the hidden-state history during training. It treats early iterations as detached 'warm-up' steps and applies loss only at the final step. This approach mitigates the biases towards greedy intermediate behaviors and reduces wall-clock costs associated with training.