Open Brain
arXivPaperNeeds Review

Is Persona Generalization Predictable?

Fine-tuning choices can shift a model’s behavior beyond the target domain. Layer selection may reduce unsafe transfer while preserving in-domain learning, a useful distinction when adapting models.

arXiv · Nov 9, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Across Qwen and Llama models, in-domain and out-of-domain persona rates followed a shared, approximately probit-linear relationship. Learning rate, weight decay, and training duration mostly moved models along that line; changing the tuned layers could change its slope.

Practical Implication

When adapting a model, track behavior outside the training domain as well as target performance. For unsafe behavior, tuning only later layers reduced transfer while retaining stronger in-domain learning.

Agent-Ready Context
Across Qwen and Llama models, in-domain and out-of-domain persona rates followed a shared, approximately probit-linear relationship. Learning rate, weight decay, and training duration mostly moved models along that line; changing the tuned layers could change its slope.

When adapting a model, track behavior outside the training domain as well as target performance. For unsafe behavior, tuning only later layers reduced transfer while retaining stronger in-domain learning.

The result describes patterns across the settings studied; it does not establish that later-layer tuning will limit unwanted transfer for every model or task.
Connected Context · Feed7 Judgment

This suggests persona generalization follows a shared in-domain/out-of-domain pattern across the studied Qwen and Llama models, with tuning choices mostly shifting performance along that pattern. Which layers are tuned can change the trade-off, so later-layer tuning is a tested way to reduce unsafe transfer, not a universal guarantee.

Context Map
modelresearch
Uncertainty
The result describes patterns across the settings studied; it does not establish that later-layer tuning will limit unwanted transfer for every model or task.