Pointer Value Retrieval (PVR)

Pointer Value Retrieval (PVR) 是一个用于研究人工神经网络推理能力极限的新基准。PVR 基于间接推理，这是人类智能的一个标志，其中第一阶段（任务）包含解决第二阶段（任务）的指令。在 PVR 中，这通过任务输入的一部分作为指针，给出不同输入位置的指令，形成输出。这种使用间接推理的方式允许通过增加功能复杂性和分布偏移来系统地改变任务难度，从而进行细致的测试，揭示神经网络泛化的细微失败和意外成功，表明在这个基准上有许多有前途的探索方向。

Pointer Value Retrieval (PVR) is a novel benchmark developed to explore the extreme limits of the reasoning capabilities of artificial neural networks. PVR is founded on indirect reasoning, a defining hallmark of human intelligence, wherein the first task contains instructions for solving the second task. Within the PVR framework, this is realized by utilizing a portion of the task input as a pointer that issues directives referring to distinct input positions to generate the final output. This approach of employing indirect reasoning enables the systematic adjustment of task difficulty by escalating functional complexity and distribution shift, facilitating fine-grained testing that uncovers subtle failures and unexpected successes in neural network generalization, and pointing to a multitude of promising research directions for this benchmark.