核心要点
- 问题/背景
- 这篇论文提出 GRAM，把 recursive latent reasoning 从确定性单轨迹更新推进到概率生成式多轨迹计算。
- 方法/机制
- 它承认并继承 HRM、TRM、Looped Transformers 这类 RRM 的核心思路：通过共享 transition functions 反复 refine persistent latent state，用递归深度替代更长 token trace 或更大参数量。
- 结果/证据
- GRAM 的关键变化是把每一步 latent transition 变成随机潜变量过程，并用 amortized variational inference 训练，从而能同时沿递归深度和并行采样宽度做 inference-time scaling。
- 收录价值
- 它值得收录，因为它为 reasoning model 提供了介于 CoT、自回归 token 扩展和 deterministic HRM/TRM 之间的新计算范式：多假设、可采样、可生成的 latent recursive reasoning。
原始摘要与中文对照
中文对照翻译
未来的神经推理系统应如何实现扩展计算？递归推理模型（RRMs）通过使用共享转换函数执行迭代潜在状态细化，为自回归序列扩展提供了一种有前景的替代方案。然而，现有的RRMs大多是确定性的，遵循单一潜在轨迹并收敛于单一预测。我们引入了生成式递归推理模型（GRAM），这是一个将递归潜在推理转化为概率性多轨迹计算的框架。GRAM将推理建模为随机潜在轨迹，通过递归深度和并行轨迹采样，实现多重假设、替代解决方案策略以及推理时扩展。这产生了一个支持通过pθ (y | x)进行条件推理，以及在输入固定或缺失时通过pθ (x)进行无条件生成的潜在变量生成模型。GRAM通过摊销变分推断进行训练，在结构化推理和多解决方案约束满足任务上优于确定性循环和递归基线，同时展示了无条件生成能力。https://ahn-ml.github.io/gram-website
原始摘要
How should future neural reasoning systems implement extended computation? Recursive Reasoning Models (RRMs) offer a promising alternative to autoregressive sequence extension by performing iterative latent-state refinement with shared transition functions. Yet existing RRMs are largely deterministic, following a single latent trajectory and converging to a single prediction. We introduce Generative Recursive reAsoning Models (GRAM), a framework that turns recursive latent reasoning into probabilistic multi-trajectory computation. GRAM models reasoning as a stochastic latent trajectory, enabling multiple hypotheses, alternative solution strategies, and inference-time scaling through both recursive depth and parallel trajectory sampling. This yields a latent-variable generative model supporting conditional reasoning via pθ (y | x) and, with fixed or absent inputs, unconditional generation via pθ (x). Trained with amortized variational inference, GRAM improves over deterministic recurrent and recursive baselines on structured reasoning and multi-solution constraint satisfaction tasks, while demonstrating an unconditional generation capability. https://ahn-ml.github.io/gram-website