Language model behavior transfers between domains on a predictable line; later-layer fine-tuning reduces slope

Across model families and hyperparameters, out-of-domain behavior rates are linearly predictable from in-domain rates along a single probit-linear line; fine-tuning only later layers lowers the line's slope for unsafe behaviors without sacrificing in-domain performance.

Independent

Olina Mukherjee · Arnav Goel · Aditi Raghunathan · Jacob Mitchell Springer

Research Digest·11 November 2026·2 min read

In one line

Persona generalization across domains is predictably probit-linear; fine-tuning later layers reduces unsafe transfer.

Before you read this

An idea this paper builds on

activation steeringActivation steering modifies a model's internal vector representations during inference to shift behavior, such as editing outputs, removing capabilities, or controlling trust.72 papers →

Read it as

When you give a language model extra training (fine-tuning) in one subject, it can start acting differently in other subjects, possibly in unsafe ways. This is persona generalization, which is hard to predict.

Researchers found a pattern: the chance of the new behavior in the training subject and in other subjects follows a shared curve for different models and settings. They also found that fine-tuning only the model's later layers reduces unwanted spread while keeping the training performance.

This means you can predict out-of-subject behavior from in-subject measures. For safety, limiting updates to deeper layers is a practical way to control harmful spillover.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

persona generalization on-the-line phenomenon probit-linear relationship supervised fine-tuning activation steering low-rank adaptation