MELD: Multi-Task Detector for AI-Generated Text
Researchers have introduced MELD (Multi-Task Equilibrated Learning Detector), a new detector for AI-generated text designed to be robust against attacks and adversarial rewrites. Unlike existing detectors that optimize a single AI/Human objective, MELD enriches binary detection with auxiliary supervision by attaching generator-family, attack-type, and source-domain heads to a shared encoder. It balances four losses using learned weights. The approach aims to maintain low false-positive rates and transfer to unseen generators and domains. The paper is available on arXiv under ID 2605.06903.
Key facts
- MELD stands for Multi-Task Equilibrated Learning Detector
- It is designed for AI-generated text detection
- It uses auxiliary supervision with generator-family, attack-type, and source-domain heads
- It balances four losses with learned weights
- Aims to be robust to attacks and adversarial rewrites
- Transfers to unseen generators and domains
- Operates at low false-positive rates
- Paper available on arXiv: 2605.06903
Entities
Institutions
- arXiv