Model Editing for New Document Integration in Generative Information Retrieval

A research team led by Zhen Zhang has introduced DOME, a breakthrough 'model editing' technique designed to solve a critical scalability issue in next-generation AI search. Generative retrieval (GR) models, which work by generating document identifiers (docIDs) in response to queries, traditionally struggle to incorporate new information. The standard solution—full retraining—is computationally expensive, resource-intensive, and causes 'catastrophic forgetting,' where the model loses knowledge of older documents. DOME directly edits a model's parameters to learn the correct mapping for new docIDs, bypassing the need for retraining the entire system and enabling practical, real-time updates.

The core innovation is a hybrid-label adaptive training strategy that creates highly discriminative 'edit vectors.' This method combines soft labels (preserving query semantics) with hard labels (enforcing precise docID mapping) to make updates distinct and effective. In experiments on major benchmarks like NQ and MS MARCO, DOME successfully integrated new documents while maintaining performance on the original collection. Crucially, it achieved this with only about 60% of the training time required by traditional incremental training. This efficiency breakthrough addresses the primary bottleneck preventing GR models from scaling, paving the way for more dynamic, up-to-date, and practical AI-powered search engines that can learn continuously without degrading.