colbert

This paper introduces MoganColBERT-TR, a late-interaction multi-vector retrieval model for Turkish, which achieves competitive zero-shot performance on Turkish BEIR datasets through distillation training from previous encoders.

The author expresses excitement about the immediate high-quality results from ColBERT models in machine learning and hints at upcoming open model releases, demonstrating the efficiency of late interaction technology without storage overhead.

Turbopuffer announces beta support for late interaction, enabling models like ColBERT to represent text as token-level vectors, combining a fast single-vector ANN first pass with exact late interaction reranking to improve recall.

Liquid AI demonstrates using LFM2.5-ColBERT-350M as a filter to select only the five most relevant tools from 151 options, reducing latency and improving tool selection accuracy.

LFM2.5-ColBERT-350M is a model that reliably selects the most relevant tools from a set of 151, saving tokens and improving accuracy, ideal for agentic edge models.

Liquid AI introduces LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M, two multilingual retrieval models optimized for fast and accurate search across 11 languages, with latency as low as 1.5ms.

The authors found that regularizing ColBERT models fixes the efficient ANN methods MUVERA and SMVE, which had broken on modern ColBERT models, simplifying multi-vector retrieval infrastructure in an unexpected way.

Tony Wu released late-interaction-kernels (LIK): fused Triton kernels for MaxSim, the scoring step behind ColBERT and ColPali, integrated into PyLate and colpali-engine, offering memory efficiency and performance gains.

ColBERTSaR proposes an embedding quantization method using product quantization to transform ColBERT's index into a true inverted index, reducing index size by 50-70% compared to one-bit PLAID while preserving retrieval effectiveness.

LateOn, a new generation ColBERT model, achieves a nearly 10-point improvement over v2 on BEIR and generalizes well outside BEIR, with the same usage in PyLate.

This paper proposes Single-stage Sparse Retrieval (SSR), which replaces K-means clustering with sparse autoencoders and inverted indexing, achieving 15x faster indexing and halved retrieval latency while improving accuracy on the BEIR benchmark.

LiquidAI releases LFM2.5-ColBERT-350M, a late-interaction multilingual retrieval model, along with a dense bi-encoder variant, both built on LFM2.5-350M-Base, supporting 11 languages and designed as drop-in replacements for RAG pipelines.

暗示即将以低调方式发布一个强大的开源多语言ColBERT搜索模型。

Perplexity AI releases pplx-embed-v1-late-0.6b, a small ColBERT late-interaction embedding model for retrieval, fine-tuned from their existing embedding model and optimized for MaxSim scoring, now open-source on HuggingFace.

Releases a kernel on Hugging Face that accelerates MaxSim late-interaction retrieval by using tiled scoring with SIMD group matrix operations (Metal and WMMA), achieving 3–5× speedup over the naive implementation.

The article highlights how ColBERT models, despite being smaller and older, outperform larger models like Qwen3-embed-8B when coupled with late interaction techniques and minimal fine-tuning.

Reason-ModernColBERT achieves near-perfect results on BrowseComp-Plus, surpassing SOTA and models 54× larger, then Agent-ModernColBERT further improves with minimal training.

Raphael released two open-source retrieval models, LateOn (ColBERT multi-vector) and DenseOn (single-vector), each 149M parameters and outperforming 4× larger models on BEIR.

A keynote recording argues that late interaction retrieval (e.g., ColBERT-style) is the most promising direction in AI-scale information retrieval research, contending that single-vector dense retrieval is fundamentally flawed and that the IR community must raise its ambitions significantly. The talk introduces the LIMIT benchmark as evidence of dense retrieval's generalization failures and calls for a paradigm shift by 2030.