Zhyper: Factorized Hypernetworks for Conditioned LLM Fine-Tuning
This addresses the need for parameter-efficient fine-tuning of LLMs for cultural alignment and semantic conditioning, offering an incremental improvement over prior LoRA-based methods.
The paper tackles the problem of conditioning large language models (LLMs) on specific cultural or semantic norms by proposing Zhyper, a factorized hypernetwork framework that generates context-aware LoRA adapters from text, achieving competitive performance with up to 26x fewer parameters than state-of-the-art methods.
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.