Few-shot Style-Conditioned LLM Text Generation via Latent Interpolation

Moemen Gaafar, Anurag Sarkar, Matthew Guzdial

Keywords: style-conditioned text generation, few-shot, transfer learning, style representation

TL;DR: We learn a latent space representing different finetuned styles and can generate new style-conditioned LLMs with only a few samples

Abstract: We propose a novel, model-agnostic approach for adapting large language models (LLMs) in a few-shot manner to arbitrary styles using text samples from a given author. Rather than use predefined features, our method defines style in terms of LLM model weights and uses a variational autoencoder (VAE) to construct a latent space of these weights, allowing for a generic style representation. Our approach leverages interpolation in this latent embedding space of model weights to generate novel fine-tuned models for low-resource authors. We evaluate this approach compared to reported results, finetuning, and prompting across three datasets. Results indicate that our method outperforms our baselines in low-resource settings.

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12757