HyperPEFTNet: Parameter-EfficientHypernetworks for Persona Synthesis

Personalizing transformer language models to reproduce stable user writing stylesat population scale is challenging because standard pipelines store and load aseparate fine-tuned model or adapter for each user. We present HyperPEFTNet,a trait-conditioned hypernetwork that generates parameter-efficient fine-tuning(PEFT) weights for a frozen decoder from a compact author descriptor derivedfrom stylometric and activity aggregates. The same descriptor-to-weights inter-face supports multiple PEFT families—including low-rank adaptation, bottleneckadapters, bias-only tuning, and prefix tuning—enabling checkpoint-free personaswitching without per-user adapter storage. We train on multi-speaker Redditthreads from 10,000 highly active users and evaluate three complementary aspects:predictive fit on held-out replies under teacher forcing, controlled representationprobes that test whether persona-relevant cohort labels remain recoverable underneutral prompts, and deployment costs such as per-user storage and persona-swapoverhead. Across several PEFT heads, descriptor-conditioned weight generationimproves teacher-forced likelihood relative to shared PEFT baselines. In repre-sentation probes, cohort structure is stronger when conditioning is intact andcollapses under descriptor-shuffle and zero-descriptor controls, indicating thatcompact user descriptors meaningfully steer internal representations. Becauseper-user state is a small numeric vector, HyperPEFTNet keeps per-user storageconstant and supports scalable, retrieval-free persona conditioning with a singleshared model.