Node segmentation

Reasoning traces are segmented into nodes, which are non-overlapping, semantically atomic segments of text. Each node is assigned a distinct label based on its semantic role.

Split into sentences by default.

Split sentences into smaller units when:

Do not split sentences into smaller units when:

The segmented nodes should be (1) fully covering the original text, and (2) not overlapping with each other. Whitespaces between two nodes are appended to the previous node, ensuring that every node starts with a non-whitespace character.

Node labels

$\color{FFADAD}{■}$ Planning

Planning introduces the content of the following nodes. It can be both coarse, high-level direction that affects tens to hundreds of nodes, or highly local plans that only affect the next node.

The rule of thumb for finding a Planning node is to check if it includes active verbs like find, recall, check, solve, or replace, with first-person subjects (I, me).