Zhyper: Factorized Hypernetworks for Conditioned LLM Fine-Tuning

TL;DR

Zhyper introduces a parameter-efficient hypernetwork approach to fine-tune large language models for conditioned generation, achieving competitive results with significantly fewer parameters and better out-of-domain generalization.

Contribution

The paper presents Zhyper, a novel factorized hypernetwork framework that generates context-aware adapters from textual descriptions, reducing parameter count while enhancing conditioning capabilities.

Findings

Achieves competitive performance with up to 26x fewer parameters.

Improves cultural alignment and out-of-domain generalization.

Effectively captures fine-grained contextual values.

Abstract

Large Language Model (LLM) conditioning refers to instructing an LLM to generate content in accordance with the norms and values of a specific culture, beliefs of a particular political orientation, or any desired text-specified semantic conditioning. Unfortunately, prompt engineering does not ensure that LLMs behave in accordance with a desired conditioning due to the inductive bias of the pre-training and alignment datasets. Prior works have focused on fine-tuning LLMs by directly conditioning the LoRA weights; however, such methods introduce a large number of parameters. As a remedy, we propose Zhyper, a parameter-efficient factorized hypernetwork framework that generates context-aware LoRA adapters from textual descriptions. Experiments on multiple benchmarks show that Zhyper achieves competitive performance with up to 26x fewer parameters than the state-of-the-art baselines. Furthermore, we extend Zhyper to cultural alignment, demonstrating improved generalization to out-of-domain settings and a better capturing of fine-grained contextual values.