# Preprint demonstrates language-model behavior edits without fine-tuning

_Published Friday, October 16, 2026 at 8:12 AM EDT · Science, AI · Latest · Tier 2 — Notable_

Researchers report a method for changing language-model behavior without fine-tuning, using statistics collected during forward passes to construct edits to model weights. Their Training-Free Task Vectors method translates activation steering vectors into rank-one weight changes.

The preprint reports that these edits amplify, suppress and combine target behaviors while preserving general knowledge and problem-solving skills. Addition combines learned changes, while subtraction removes them. In evaluations against other editing and steering methods, the researchers report stronger trait control with better or competitive preservation of utility; the experiments concern behavioral control tasks.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2609.09054)

---
Canonical: https://techandbusiness.org/newswire/Xtg1mD6NTYQVh74Asnc3SR
Published: 2026-10-16T12:12:57.286Z
Story chronology: 2026-10-16T04:00:00.000Z
Retrieved: 2026-10-16T15:28:08.119Z
Publisher: Tech & Business (techandbusiness.org)