Query Drift Compensation: Enabling Compatibility in Continual Learning of Retrieval Embedding Models

2025/05/27 by Dipam Goswami, Liying Wang, Goswami, Dipam +5 · 1 citation
Computer Science · #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Information Retrieval (cs.IR) #Machine Learning (cs.LG) #Machine Learning and Algorithms #Machine Learning and ELM

paper · pdf · doi:10.48550/arxiv.2506.00037

openalex publication_date 2025/05/27 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Text embedding models enable semantic search, powering several NLP applications like Retrieval Augmented Generation by efficient information retrieval (IR). However, text embedding models are commonly studied in scenarios where the training data is static, thus limiting its applications to dynamic scenarios where new training data emerges over time. IR methods generally encode a huge corpus of documents to low-dimensional embeddings and store them in a database index. During retrieval, a semantic search over the corpus is performed and the document whose embedding is most similar to the query embedding is returned. When updating an embedding model with new training data, using the already indexed corpus is suboptimal due to the non-compatibility issue, since the model which was used to obtain the embeddings of the corpus has changed. While re-indexing of old corpus documents using the updated model enables compatibility, it requires much higher computation and time. Thus, it is critical to study how the already indexed corpus can still be effectively used without the need of re-indexing. In this work, we establish a continual learning benchmark with large-scale datasets and continually train dense retrieval embedding models on query-document pairs from new datasets in each task and observe forgetting on old tasks due to significant drift of embeddings. We employ embedding distillation on both query and document embeddings to maintain stability and propose a novel query drift compensation method during retrieval to project new model query embeddings to the old embedding space. This enables compatibility with previously indexed corpus embeddings extracted using the old model and thus reduces the forgetting. We show that the proposed method significantly improves performance without any re-indexing. Code is available at https://github.com/dipamgoswami/QDC.

Citations

• Exemplar-free Continual Representation Learning via Learnable Drift Compensation
• Resurrecting Old Classes with New Data for Exemplar-Free Continual Learning
• Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements
• Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models
• Elastic Feature Consolidation for Cold Start Exemplar-Free Incremental Learning
• Discovery architecture for cultural heritage: layered retrieval, institutional authority, and the limits of keyword search
• Jina Embeddings 2: 8192-Token General-Purpose Text Embeddings for Long Documents
• FeCAM: Exploiting the Heterogeneity of Class Distributions in Exemplar-Free Continual Learning
• L²R: Lifelong Learning for First-stage Retrieval with Backward-Compatible Representations
• Advancing continual lifelong learning in neural information retrieval: definition, dataset, framework, and empirical evaluation
• Towards General Text Embeddings with Multi-stage Contrastive Learning
• IncDSI: Incrementally Updatable Document Retrieval
• Open-World Continual Learning: Unifying Novelty Detection and Continual Learning
• A Comprehensive Survey of Continual Learning: Theory, Method and Application
• In-Context Retrieval-Augmented Language Models
• DSI++: Updating Transformer Memory with New Documents
• Text Embeddings by Weakly-Supervised Contrastive Pre-training
• FeTrIL: Feature Translation for Exemplar-Free Class-Incremental Learning
• NEVIS'22: A Stream of 100 Tasks Sampled from 30 Years of Computer Vision Research
• MTEB: Massive Text Embedding Benchmark
• Atlas: Few-shot Learning with Retrieval Augmented Language Models
• Continual Learning for Visual Search with Backward Consistent Feature Embedding
• A Closer Look at Rehearsal-Free Continual Learning
• Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning
• SGPT: GPT Sentence Embeddings for Semantic Search
• Continual Learning of Long Topic Sequences in Neural Information Retrieval
• Unsupervised Dense Information Retrieval with Contrastive Learning
• Forward Compatible Training for Large-Scale Embedding Retrieval Systems
• BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models
• Continual learning in cross-modal retrieval
• Studying Catastrophic Forgetting in Neural Ranking Models
• Pretrained Transformers for Text Ranking: BERT and Beyond
• Embedding-based Zero-shot Retrieval through Query Generation
• Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• Sparse, Dense, and Attentional Representations for Text Retrieval
• ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT
• Dense Passage Retrieval for Open-Domain Question Answering
• Semantic Drift Compensation for Class-Incremental Learning
• Towards Backward-Compatible Representation Learning
• Three scenarios for continual learning
• End-to-End Retrieval in Continuous Space
• HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering
• Representation Learning with Contrastive Predictive Coding
• FEVER: a large-scale dataset for Fact Extraction and VERification
• Overcoming catastrophic forgetting with hard attention to the task
• PackNet: Adding Multiple Tasks to a Single Network by Iterative Pruning
• Measuring Catastrophic Forgetting in Neural Networks
• Attention Is All You Need
• iCaRL: Incremental Classifier and Representation Learning
• MS MARCO: A Human Generated MAchine Reading COmprehension Dataset
• Learning without Forgetting
• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

Cited by

• Online-Optimized RAG for Tool Use and Function Calling

Related

• Conventional Contrastive Learning Often Falls Short: Improving Dense Retrieval with Cross-Encoder Listwise Distillation and Synthetic Data
• Augmented Embeddings for Custom Retrievals
• DSI++: Updating Transformer Memory with New Documents
• REFINE on Scarce Data: Retrieval Enhancement through Fine-Tuning via Model Fusion of Embedding Models
• Towards Robust Text Retrieval with Progressive Learning