LLM-HYPER Deploys LLM Hypernetworks for Training-Free CTR Generation in Cold-Start Ads
LLM-HYPER generates CTR model weights via LLM hypernetworks and multimodal prompting, outperforming baselines by 55.9% NDCG@10 and deployed in production to shorten cold-start ad personalization.
Researchers from a top U.S. e-commerce platform introduced LLM-HYPER, which treats LLMs as hypernetworks to generate parameters for a linear CTR estimator directly from multimodal ad content using few-shot Chain-of-Thought prompting (Ma et al., arXiv:2604.12096).
The method retrieves similar past campaigns via CLIP embeddings, reasons over customer intent and feature influence, then applies normalization and calibration to align outputs with production CTR distributions, reporting 55.9% NDCG@10 gains versus cold-start baselines and positive A/B test results leading to deployment (Ma et al., arXiv:2604.12096; Ha et al., arXiv:1609.09106).
LLM-HYPER extends hypernetwork foundations (Ha et al., arXiv:1609.09106) and CTR architectures such as DCN (Wang et al., arXiv:1708.05123), differing from meta-learning or proxy-feature cold-start techniques by synthesizing weights without gradient updates; prior coverage overlooked the production integration that compresses the feedback loop for new promotional ads on high-revenue platforms.
Industrial patterns show generative models entering core prediction stacks at scale, consistent with unreported shifts at other ad platforms where LLM-derived components reduce ramp-up from weeks to immediate serving.
AXIOM: LLM-HYPER shows LLMs generating full CTR model parameters from ad creatives alone, letting new promotions bypass data collection and integrate directly into production ad systems at major ecommerce sites.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2604.12096)
- [2]HyperNetworks(https://arxiv.org/abs/1609.09106)
- [3]Deep & Cross Network for Ad Click Predictions(https://arxiv.org/abs/1708.05123)