How Does Prefix Matter in Reasoning Model Tuning?
Raj Vardhan Tomar, Preslav Nakov, Yuxia Wang · 2026
Read How Does Prefix Matter in Reasoning Model Tuning? by Raj Vardhan Tomar, Preslav Nakov, Yuxia Wang (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-2601-01624.txt.
Description
Tomar, Nakov, and Wang ask whether opening sentences that many trainers delete actually help reasoning models. Fine-tuning data often begin with boilerplate lines about safety or careful reasoning, and common practice strips them out. The team instead test whether those prefixes act as light alignment signals that steer answers toward safer, clearer paths.
They fine-tuned three models from one reasoning series while varying how many training examples kept prefixes, from none to all. Safety scores rose by up to about six percent on hard test sets, and grade-school math rose by about seven percent. Factual and coding tasks gained little or slipped, suggesting prefixes help most for structured reasoning.
Catalog record
| CREATOR | Raj Vardhan Tomar, Preslav Nakov, Yuxia Wang |
|---|---|
| YEAR | 2026 |
| 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/2601-01624