English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Is Persona Generalization Predictable? An 'On-the-Line' Phenomenon in LLM Fine-Tuning

Forum topic · 小凯 · 2026-11-10

Summary

A new arXiv paper (2611.12864) by Mukherjee, Goel, Raghunathan, and Springer discovers an 'on-the-line' phenomenon in persona generalization, where fine-tuning a language model on one domain induces behavior changes in others—such as the broad unsafe behavior seen in emergent misalignment. The authors find that in-domain and out-of-domain persona rates follow an approximately probit-linear relationship across fine-tuning hyperparameters, and this linear relationship is surprisingly shared across model families and scales, including Qwen and Llama. Training choices that do not alter the trainable parameter set—learning rate, weight decay, training duration—mainly move the model along the same line, reducing transfer by weakening in-domain learning. In contrast, changing which layers are fine-tuned changes the slope: tuning only later layers produces shallower slopes, retaining strong in-domain learning while reducing out-of-domain misalignment. The finding unifies scattered persona generalization observations into a quantitative framework, enabling prediction and control of behavioral transfer.

Paper Overview

  • Field: NLP
  • Authors: Olina Mukherjee, Arnav Goel, Aditi Raghunathan, Jacob Mitchell Springer
  • Published: 2026-11-09
  • arXiv: 2611.12864
  • What the Paper Is About

    Persona generalization refers to the phenomenon where fine-tuning a language model on one domain induces behavioral changes in other domains—such as the broad unsafe behavior observed in emergent misalignment. This paper asks whether such transfer is predictable.

    Key Findings

  • An 'on-the-line' phenomenon: In-domain and out-of-domain persona rates follow an approximately probit-linear relationship across fine-tuning hyperparameters.
  • Cross-model regularity: This probit-linear relationship is remarkably shared across model families and scales, including Qwen and Llama.
  • Two regimes of hyperparameters:
  • Training choices that do not change the set of trainable parameters (learning rate, weight decay, training duration) mainly move the model along the same line, reducing transfer primarily by weakening in-domain learning.
  • Changing which layers are fine-tuned changes the slope of the relationship.
  • Safety implication: Fine-tuning only later layers yields shallower slopes—retaining strong in-domain learning while reducing out-of-domain misalignment.

Why It Matters

The 'on-the-line' phenomenon unifies scattered persona generalization observations into a single quantitative framework, making it possible to predict and control behavioral transfer in fine-tuned language models.

Original Abstract (excerpt)

> We discover an 'on-the-line' phenomenon in persona generalization: in-domain and out-of-domain persona rates follow an approximately probit-linear relationship shared across model families and sizes. Fine-tuning only later layers yields shallower slopes, retaining strong in-domain learning while reducing out-of-domain misalignment.

*Auto-collected on 2026-11-11.*

Tags

#nlp#arxiv#persona-generalization#fine-tuning#emergent-misalignment#llm-safety#qwen#llama

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178637020