RTNewLine (Paragraph-based segmentation via blank lines)

Idea

RTNewLine is a minimal rule-based baseline that segments a reasoning trace using formatting structure rather than lexical or model-based cues.

It splits the trace into segments at blank-line boundaries, i.e., occurrences of \n\n, which often correspond to paragraph breaks, step breaks, or deliberate spacing inserted by the generator or annotator.

This engine is particularly useful as:

• a fast baseline,
• a preprocessing heuristic when traces already contain meaningful line breaks,
• a robust fallback in environments without model dependencies.

---

Method (high-level)

1. Find segment start positions Identify all positions in the trace immediately following either:
   • the start of the string (\A), or
   • a blank-line delimiter (\n\n)

   Concretely, the engine collects:

   • positions = [m.end() for m in re.finditer(r'\n\n|\A', trace)]
2. Create consecutive spans Pair each start position with the next start position (or end of trace for the last segment):
   • offsets = zip(positions, positions[1:] + [len(trace)])
3. Return offsets Emit character offsets and assign "UNK" labels to all segments.

---

Models used

None. This engine is purely regex-based.

---

Parameters

RTNewLine does not require any configuration parameters. Any **kwargs are ignored.

---

Usage

from rt_seg import RTSeg
from rt_seg import RTNewLine

trace = "Step 1: ...\n\nStep 2: ...\n\nFinal: ..."

segmentor = RTSeg(engines=RTNewLine)
offsets, labels = segmentor(trace)

segments = [trace[s:e] for s, e in offsets]
for seg in segments:
    print("---")
    print(seg)