Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CREAM: Continual Retrieval on Dynamic Streaming Corpora with Adaptive Soft Memory

HuiJeong Son, Hyeongu Kang, Sunho Kim, Subeen Ho, SeongKu Kang, Dongha Lee, Susik Yoon

TL;DR

CREAM tackles continual information retrieval in dynamic streaming corpora where topic distributions shift over time. It introduces adaptive soft memory and three mechanisms—fine-grained token-level similarity, regularized cluster prototypes via RP-LSH, and stratified coreset sampling—to enable fully unsupervised continual learning without ground-truth labels. Empirical results on LoTTE and MSMARCO show substantial improvements over unsupervised baselines and competitive performance with supervised methods, highlighting strong practical potential for real-time retrieval systems. By preserving past knowledge while adapting to new topics, CREAM reduces forgetting and computational overhead in streaming IR deployments.

Abstract

Information retrieval (IR) in dynamic data streams is emerging as a challenging task, as shifts in data distribution degrade the performance of AI-powered IR systems. To mitigate this issue, memory-based continual learning has been widely adopted for IR. However, existing methods rely on a fixed set of queries with ground-truth relevant documents, which limits generalization to unseen queries and documents, making them impractical for real-world applications. To enable more effective learning with unseen topics of a new corpus without ground-truth labels, we propose CREAM, a self-supervised framework for memory-based continual retrieval. CREAM captures the evolving semantics of streaming queries and documents into dynamically structured soft memory and leverages it to adapt to both seen and unseen topics in an unsupervised setting. We realize this through three key techniques: fine-grained similarity estimation, regularized cluster prototyping, and stratified coreset sampling. Experiments on two benchmark datasets demonstrate that CREAM exhibits superior adaptability and retrieval accuracy, outperforming the strongest method in a label-free setting by 27.79\% in Success@5 and 44.5\% in Recall@10 on average, and achieving performance comparable to or even exceeding that of supervised methods.

CREAM: Continual Retrieval on Dynamic Streaming Corpora with Adaptive Soft Memory

TL;DR

CREAM tackles continual information retrieval in dynamic streaming corpora where topic distributions shift over time. It introduces adaptive soft memory and three mechanisms—fine-grained token-level similarity, regularized cluster prototypes via RP-LSH, and stratified coreset sampling—to enable fully unsupervised continual learning without ground-truth labels. Empirical results on LoTTE and MSMARCO show substantial improvements over unsupervised baselines and competitive performance with supervised methods, highlighting strong practical potential for real-time retrieval systems. By preserving past knowledge while adapting to new topics, CREAM reduces forgetting and computational overhead in streaming IR deployments.

Abstract

Information retrieval (IR) in dynamic data streams is emerging as a challenging task, as shifts in data distribution degrade the performance of AI-powered IR systems. To mitigate this issue, memory-based continual learning has been widely adopted for IR. However, existing methods rely on a fixed set of queries with ground-truth relevant documents, which limits generalization to unseen queries and documents, making them impractical for real-world applications. To enable more effective learning with unseen topics of a new corpus without ground-truth labels, we propose CREAM, a self-supervised framework for memory-based continual retrieval. CREAM captures the evolving semantics of streaming queries and documents into dynamically structured soft memory and leverages it to adapt to both seen and unseen topics in an unsupervised setting. We realize this through three key techniques: fine-grained similarity estimation, regularized cluster prototyping, and stratified coreset sampling. Experiments on two benchmark datasets demonstrate that CREAM exhibits superior adaptability and retrieval accuracy, outperforming the strongest method in a label-free setting by 27.79\% in Success@5 and 44.5\% in Recall@10 on average, and achieving performance comparable to or even exceeding that of supervised methods.
Paper Structure (34 sections, 1 theorem, 10 equations, 5 figures, 7 tables, 4 algorithms)

This paper contains 34 sections, 1 theorem, 10 equations, 5 figures, 7 tables, 4 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Existing Efforts
  4. Main Idea and Contributions
  5. Related Work
  6. Information Retrieval
  7. Continual Learning on IR
  8. Problem Setting
  9. Methodology
  10. Retrieval
  11. Memory Update
  12. Cluster Assignment with Regularized Prototype.
  13. Lightweight Cluster Maintenance.
  14. Encoder Update
  15. Query Selection.
  16. ...and 19 more sections

Key Result

theorem 1

(Sufficient LSH bitsize) When generating prototypes from $M$ token embeddings, the sufficient number of LSH bits is determined as $\log_2 \left( \frac{8 \ln M}{\varepsilon^2} \right)$ at the optimal distortion rate $\varepsilon = \frac{1}{3\sqrt{e}}$.

Figures (5)

  • Figure 1: Comparison of existing (top) and our (bottom) approaches for memory-based continual retrieval.
  • Figure 2: Overall framework of CREAM with three components: (1) a retrieval component that returns the relevant documents to a given query with the up-to-date encoder; (2) a memory update component that captures the recent knowledge while preserving previously acquired information through streaming clustering with regularized prototypes; and (3) an encoder update component that facilitates self-supervised training using contrastive objective and the structure of soft memory.
  • Figure 3: Training time analysis results.
  • Figure 4: Processing time analysis results.
  • Figure 5: Soft memory with query, pseudo-positive samples, and negative samples in sessions 0, 3, 6, and 9.

Theorems & Definitions (1)

  • theorem 1