Not the Example, but the Process: How Self-Generated Examples Enhance LLM Reasoning

Daehoon Gwak, Minseo Jung, Junwoo Park, Minho Park, ChaeHun Park Junha Hyung Jaegul Choo · 2026

Read Not the Example, but the Process: How Self-Generated Examples Enhance LLM Reasoning by Daehoon Gwak, Minseo Jung, Junwoo Park, Minho Park, ChaeHun Park Junha Hyung Jaegul Choo (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-15863.txt.

Description

Gwak and colleagues argue that the benefit of self-made few-shot samples comes not from the samples themselves but from the act of making them. To validate their claim, the team fully tests three prompting routes for in-context learning across thought-heavy tasks and varied language model designs.

The three routes are zero-shot prompting; joined prompting, where language models both make and solve problems in one unified prompt; and split prompting. In split prompting, self-made examples serve again as in-context examples, while the context of their creation stays out. Tests across five widely used model designs show joined prompting always beating both zero-shot and split prompting. Split prompting offers only slight gains over zero-shot. The findings suggest the edge of self-making prompting comes from the act of problem making, not the samples themselves, giving tips for building more working prompting routes.

Catalog record

CREATORDaehoon Gwak, Minseo Jung, Junwoo Park, Minho Park, ChaeHun Park Junha Hyung Jaegul Choo
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-15863

Browse the full library · About the Tower of Babel