LLMs Enhance Auto-Bidding When Integrated with Numerical Features
A new study from arXiv (2605.05833) investigates the role of Large Language Models (LLMs) in auto-bidding for real-time advertising markets. Auto-bidding policies must optimize long-term value under delivery constraints like budget and CPA. Traditional methods rely on numerical state representations that capture delivery dynamics but fail to represent high-level intent, evolving feedback, or strategic guidance. LLMs offer semantic encoding, but their integration risks numerical precision. Through systematic studies, researchers found that LLM embeddings contain bidding-relevant cues but cannot replace numerical features. Gains occur only with careful semantic-numeric integration, not naive concatenation. The findings suggest a hybrid approach for future auto-bidding systems.
Key facts
- Auto-bidding optimizes long-horizon value under budget and CPA constraints.
- Existing methods use compact numerical state representations.
- Numerical representations cannot explicitly represent high-level intent or strategic guidance.
- LLMs encode semantic information but may sacrifice numerical precision.
- LLM embeddings contain bidding-relevant cues.
- LLMs cannot replace numerical features in auto-bidding.
- Gains require careful semantic-numeric integration.
- Naive concatenation of LLM and numerical features does not yield improvements.
Entities
Institutions
- arXiv