NumColBERT: Non-Intrusive Numeracy Injection for Late-Interaction Retrieval Models
sigir26-164 · Full research paper · Haruki Fujimaki, Makoto P. Kato
Learned Sparse Retrieval Efficiency · 50% Domain-Specific and Multilingual Den · 25% Learned Sparse Retrieval · 25%
Abstract
This study addresses the challenge of improving dense retrieval performance for queries containing numerical conditions, such as ''companies with more than one billion dollars in R&D expenditure.'' Although recent research has underscored the limitations of standard models in handling numeric information across domains such as finance, e-commerce, and medicine, existing solutions typically decompose queries into textual and numerical components and score them separately using dedicated methods. These approaches intrude upon late-interaction retrieval models such as ColBERT and incur considerable challenges in deployment, latency, and maintainability. To overcome these limitations, we propose NumColBERT, an inference-time non-intrusive method that enhances numerically conditioned retrieval while preserving the original late-interaction mechanism and providing unified scoring across textual and numerical content. Because NumColBERT retains the standard ColBERT indexing and MaxSim scoring pipeline, existing optimizations and ecosystem components developed for ColBERT can be directly reused, facilitating practical deployment. NumColBERT introduces a Numerical Gating Mechanism and a Numerical Contrastive Learning objective to enable numerical conditions to contribute more effectively to retrieval within the standard ColBERT scoring mechanism. The gating mechanism dynamically amplifies the influence of tokens carrying critical numerical constraints while suppressing context-neutral mentions such as model numbers or dates. The contrastive objective explicitly shapes the embedding space to reflect numerical magnitudes and conditions, enabling numerical values to be distinguished within the shared representation space. Experimental results show that NumColBERT substantially outperforms standard fine-tuning baselines and achieves accuracy that matches or exceeds that of prior approaches that rely on separate textual and numerical scoring. These findings demonstrate the feasibility of numerically conditioned retrieval with a non-intrusive inference pipeline and present a maintainable solution for real-world deployment.
Assigned reviewers
| # | Reviewer | Match | Why (closest evidence publications in the taxonomy) | |
|---|---|---|---|---|
| 1 | Antonio Mallia | 74 | 88 Learning Passage Impacts for Inverted Indexes (2021) Learned Sparse Retrieval Efficiency 88 Improved Learned Sparse Retrieval with Corpus-Specific Vocabularies (2024) Learned Sparse Retrieval Efficiency 84 Faster Learned Sparse Retrieval with Block-Max Pruning (2024) Learned Sparse Retrieval Efficiency |
|
| 2 | Xueguang Ma | 63 | 88 LACONIC: Dense-Level Effectiveness for Scalable Sparse Retrieval via a Two-Phase Training Curriculum (2026) Learned Sparse Retrieval Efficiency 84 Gosling Grows Up: Retrieval with Learned Dense and Sparse Representations Using Anserini (2025) Learned Sparse Retrieval Efficiency 84 SLIM: Sparsified Late Interaction for Multi-Vector Retrieval with Inverted Indexes (2023) Learned Sparse Retrieval Efficiency |
|
| 3 | Xiao Wang 0043 | 58 | 84 Reproducibility, Replicability, and Insights into Dense Multi-Representation Retrieval Models: from ColBERT to Col (2023) Learned Sparse Retrieval Efficiency 79 An Inspection of the Reproducibility and Replicability of TCT-ColBERT (2022) Learned Sparse Retrieval Efficiency 57 Effective Contrastive Weighting for Dense Query Expansion (2023) Query Reformulation and Expansion |
Area chair
| Area chair | Match | |
|---|---|---|
| Xiaoyi Zeng | 41 |
Area-chair candidates (4)
| Chenyan Xiong | 41 | 15 papers | |
| Xiaoyi Zeng assigned | 41 | 11 papers | |
| Xiaoxue Zang | 37 | 4 papers | |
| Jun Xu 0001 | 36 | 2 papers |
Candidates from the panel ranked by match score
| # | Reviewer | Match | Load | Why | |
|---|---|---|---|---|---|
| 1 | Antonio Mallia assigned | 74 | 3 | 88 Learning Passage Impacts for Inverted Indexes (2021) Learned Sparse Retrieval Efficiency 88 Improved Learned Sparse Retrieval with Corpus-Specific Vocabularies (2024) Learned Sparse Retrieval Efficiency |
|
| 2 | Franco Maria Nardini | 74 | 3 | 84 Efficient Inverted Indexes for Approximate Retrieval over Learned Sparse Representations (2024) Learned Sparse Retrieval Efficiency 84 Pairing Clustered Inverted Indexes with κ-NN Graphs for Fast Approximate Retrieval over Learned Sparse Representations (2024) Learned Sparse Retrieval Efficiency |
|
| 3 | Huiqiang Jiang | 69 | 3 | 74 RetroInfer: A Vector Storage Engine for Scalable Long-Context LLM Inference (2026) Learned Sparse Retrieval Efficiency 74 SCBench: A KV Cache-Centric Analysis of Long-Context Methods (2025) Learned Sparse Retrieval Efficiency |
|
| 4 | Xueguang Ma assigned | 63 | 3 | 88 LACONIC: Dense-Level Effectiveness for Scalable Sparse Retrieval via a Two-Phase Training Curriculum (2026) Learned Sparse Retrieval Efficiency 84 Gosling Grows Up: Retrieval with Learned Dense and Sparse Representations Using Anserini (2025) Learned Sparse Retrieval Efficiency |
|
| 5 | Sebastian Bruch 0001 | 62 | 3 | 84 Forward Index Compression for Learned Sparse Retrieval (2026) Learned Sparse Retrieval Efficiency 74 Efficient Inverted Indexes for Approximate Retrieval over Learned Sparse Representations (2024) Learned Sparse Retrieval Efficiency |
|
| 6 | Xiao Wang 0043 assigned | 58 | 3 | 84 Reproducibility, Replicability, and Insights into Dense Multi-Representation Retrieval Models: from ColBERT to Col (2023) Learned Sparse Retrieval Efficiency 79 An Inspection of the Reproducibility and Replicability of TCT-ColBERT (2022) Learned Sparse Retrieval Efficiency |
|
| 7 | Sheng-Chieh Lin | 54 | 3 | 84 Pyserini: A Python Toolkit for Reproducible Information Retrieval Research with Sparse and Dense Representations (2021) Learned Sparse Retrieval Efficiency 84 SLIM: Sparsified Late Interaction for Multi-Vector Retrieval with Inverted Indexes (2023) Learned Sparse Retrieval Efficiency |
|
| 8 | Jurek Leonhardt | 50 | 3 | 74 Efficient Neural Ranking using Forward Indexes (2022) Learned Sparse Retrieval Efficiency 74 Efficient Neural Ranking Using Forward Indexes and Lightweight Encoders (2024) Learned Sparse Retrieval Efficiency |
|
| 9 | Jingtao Zhan | 49 | 3 | 84 Joint Optimization of Multi-vector Representation with Product Quantization (2022) Learned Sparse Retrieval Efficiency 56 Learning Discrete Representations via Constrained Clustering for Effective and Efficient Dense Retrieval (2022) Approximate Nearest Neighbor Search |
|
| 10 | Shengyao Zhuang | 49 | 3 | 79 LACONIC: Dense-Level Effectiveness for Scalable Sparse Retrieval via a Two-Phase Training Curriculum (2026) Learned Sparse Retrieval Efficiency 70 Layer-wise Token Compression for Efficient Document Reranking (2026) Learned Sparse Retrieval Efficiency |
|
| 11 | Xinyu Ma 0001 | 47 | 3 | 65 Probe-and-Fetch: Dynamic KV Cache Pruning for Accelerated Long-Context Inference in Web-Scale AI Search (2026) Efficient and Long-Context RAG 59 Replication and Exploration of Generative Retrieval over Dynamic Corpora (2025) Generative Retrieval |
|
| 12 | Keping Bi | 43 | 3 | 68 Reproducibility Analysis and Enhancements for Multi-aspect Dense Retriever with Aspect Learning (2024) Retrieval Toolkits and Reproducibility 63 Pre-training with Aspect-Content Text Mutual Prediction for Multi-Aspect Dense Retrieval (2023) Sparse and Dense Retrieval Encoders |
|
| 13 | Negar Arabzadeh | 42 | 3 | 60 Predicting Efficiency/Effectiveness Trade-offs for Dense vs. Sparse Retrieval Strategy Selection (2021) Efficient Retrieval and Indexing 51 Learning Query-Space Document Representations for High-Recall Retrieval (2023) Dense Bi-Encoder Training and Representation Learning |
|
| 14 | Yuchen Li 0006 | 42 | 3 | 65 Probe-and-Fetch: Dynamic KV Cache Pruning for Accelerated Long-Context Inference in Web-Scale AI Search (2026) Efficient and Long-Context RAG 44 GS2P: a generative pre-trained learning to rank model with over-parameterization for web-scale search (2024) Web-Scale Learning to Rank |
|
| 15 | Daiting Shi | 42 | 3 | 75 Pre-trained Language Model-based Retrieval and Ranking for Web Search (2023) Learned Sparse Retrieval Efficiency 72 Pre-trained Language Model based Ranking in Baidu Search (2021) Learned Sparse Retrieval Efficiency |
|
| 18 | Corby Rosset | 41 | 3 | 70 Optimizing Query Evaluations Using Reinforcement Learning for Web Search (2018) Efficient Retrieval and Indexing 65 Dodo: Dynamic Contextual Compression for Decoder-only LMs (2024) Learned Sparse Retrieval Efficiency |
|
| 19 | Ruqing Zhang 0001 | 41 | 3 | 54 B-PROP: Bootstrapped Pre-training with Representative Words Prediction for Ad-hoc Retrieval (2021) Sparse and Dense Retrieval Encoders 54 PROP: Pre-training with Representative Words Prediction for Ad-hoc Retrieval (2021) Sparse and Dense Retrieval Encoders |
|
| 20 | Zheng Zhang 0006 | 40 | 3 | 40 Enhancing Cross-Modal Retrieval via Visual-Textual Prompt Hashing (2024) Cross-Modal Hashing and Quantization 40 Cognitive multi-modal consistent hashing with flexible semantic transformation (2022) Cross-Modal Hashing and Quantization |
|
| 21 | Venktesh V | 39 | 3 | 70 Breaking the Lens of the Telescope: Online Relevance Estimation over Large Retrieval Sets (2025) Learned Sparse Retrieval Efficiency 59 TagRec++: Hierarchical Label Aware Attention Network for Question Categorization (2024) Text Embedding and Representation Learning |
|
| 22 | Bo Chen 0023 | 39 | 3 | 69 Efficiency Unleashed: Inference Acceleration for LLM-based Recommender Systems with Speculative Decoding (2025) Efficiency and Deployment 40 LLM4Tag: Automatic Tagging System for Information Retrieval via Large Language Models (2025) Domain-Specific and Structured Retrieval |
|
| 23 | Jialong Wu 0007 | 38 | 3 | 74 SCOPE: Optimizing Key-Value Cache Compression in Long-context Generation (2025) Learned Sparse Retrieval Efficiency 40 AdaRewriter: Unleashing the Power of Prompting-based Conversational Query Reformulation via Test-Time Adaptation (2025) Query Reformulation and Expansion |
|
| 24 | Meiyu Liang | 38 | 3 | 40 Deep Unsupervised Momentum Contrastive Hashing for Cross-modal Retrieval (2023) Cross-Modal Hashing and Quantization 40 Unsupervised Multimodal Graph Contrastive Semantic Anchor Space Dynamic Knowledge Distillation Network for Cross-Media Hash Retrieval (2024) Cross-Modal Hashing and Quantization |
|
| 25 | Yanzhao Zhang | 38 | 3 | 56 SSRB: Direct Natural Language Querying to Massive Heterogeneous Semi-Structured Data (2025) Domain-Specific and Structured Retrieval 46 Unsupervised Semantic Association Learning with Latent Label Inference (2021) Text Embedding and Representation Learning |
|
| 26 | Jieming Zhu | 37 | 3 | 41 EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration (2024) Item Tokenization and Semantic IDs 40 CoST: Contrastive Quantization based Semantic Tokenization for Generative Recommendation (2024) Item Tokenization and Semantic IDs |
|
| 27 | Jinpeng Wang 0002 | 37 | 3 | 40 Hugs Bring Double Benefits: Unsupervised Cross-Modal Hashing with Multi-granularity Aligned Transformers (2024) Cross-Modal Hashing and Quantization 40 Webly Supervised Deep Attentive Quantization (2021) Cross-Modal Hashing and Quantization |
|
| 28 | Peitian Zhang | 37 | 3 | 65 Hybrid Inverted Index Is a Robust Accelerator for Dense Retrieval (2023) Efficient Retrieval and Indexing 40 C-Pack: Packed Resources For General Chinese Embeddings (2024) Text Embedding and Representation Learning |
|
| 29 | Fengling Li 0001 | 37 | 3 | 40 Cross-Domain Transfer Hashing for Efficient Cross-Modal Retrieval (2024) Cross-Modal Hashing and Quantization 40 PIC-CMH: Efficient Prompt-Infused Continual Cross-Modal Hashing (2026) Cross-Modal Hashing and Quantization |
|
| 30 | Yupeng Hou | 37 | 3 | 40 Large Language Models are Zero-Shot Rankers for Recommender Systems (2024) Efficiency and Deployment 40 Towards Large Generative Recommendation: A Tokenization Perspective (2025) Item Tokenization and Semantic IDs |
|
| 32 | Hao Deng 0011 | 37 | 3 | 40 REG4Rec: Reasoning-Enhanced Generative Model for Large-Scale Recommendation Systems (2026) Item Tokenization and Semantic IDs 38 CSMF: Cascaded Selective Mask Fine-Tuning for Multi-Objective Embedding-Based Retrieval (2025) Multi-Task and Multi-Objective Optimization |
|
| 33 | Qijiong Liu | 37 | 3 | 40 CoST: Contrastive Quantization based Semantic Tokenization for Generative Recommendation (2024) Item Tokenization and Semantic IDs 40 Learning Multi-aspect Item Palette: A Semantic Tokenization Framework for Generative Recommendation (2026) Item Tokenization and Semantic IDs |