Persona Adapters

This repository contains the persona adapter models presented in the paper Stability Regimes for Framing-Sensitive Fine-Tuning in Language Models (NeurIPS 2026).

Model Information

All adapters in this repository are LoRA adapters trained on top of the Llama-3.1-8B-Instruct base model using 15k training samples. For each framing dimension (authoritative, consensus, emotional, prestige, and sensationalist), we release three independently trained adapters. The baseline adapters were trained with alpha = 1, while all other adapters were trained with alpha = 0.3.

Within the paper framework, these framing-sensitive agent personas are constructed by fine-tuning language models with framing-conditioned loss attenuation, inducing targeted biases while preserving overall task competence. For full details, see the accompanying paper.

Resources

Code Repository: GitHub - infosenselab/frameref
Dataset: infosense/frameref
Paper: arXiv, Hugging Face

Citing this work.

Please cite the following paper when referencing this work.

@misc{de_lima_stability_2026,
author = {De Lima, Victor and Liu, Jiqun and Yang, Grace Hui},
doi = {10.48550/arXiv.2602.15273},
title = {Stability {Regimes} for {Framing}-{Sensitive} {Fine}-{Tuning} in {Language} {Models}},
url = {https://doi.org/10.48550/arXiv.2602.15273},
year = {2026}
}