Effects of structure on reasoning in instance-level Self-Discover
Sachith Gunasekara, Yasiru Ratnayake · 2025
Read Effects of structure on reasoning in instance-level Self-Discover by Sachith Gunasekara, Yasiru Ratnayake (2025) 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-2507-03347.txt.
Description
Sachith Gunasekara and Yasiru Ratnayake compare structured and free form reasoning in language models in 2025. Many systems force models to output neat structured plans. That aids trust and control, but may cost skill. The pair adapt a self discovery method to work case by case. Models build their own plan for each task in one of two forms.
Tests span many benchmarks with top open models. Free form plans won in a steady way. On a hard math test, the gain reached near a fifth over structured plans. Free form with no examples even beat structured plans with five examples. The best plan grain, case or task, depended on context. The authors urge a rethink of forced formats.
Catalog record
| CREATOR | Sachith Gunasekara, Yasiru Ratnayake |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-03347