Institution profile

Elasticsearch

Industry researchnorthamerica · us
Official website
Research library10linked papers
Opportunities0open roles
Selected work

Representative Papers

What Transfers from a VLM Teacher? Comparing Supervision Signals for Visual Document Retrieval

Oct 06, 2026

This study addresses the problem of false negative misclassification caused by incomplete annotations in visual document retrieval. We propose a hard negative discrimination and distillation framework leveraging teacher signals from Vision-Language Models (VLMs). By scoring candidate negatives with a VLM, we systematically compare knowledge transfer effects under different supervision signals and optimize the retrieval training strategy through contrastive learning and attention mechanisms. This work is the first to demonstrate that VLM-guided hard negative discrimination significantly outperforms conventional positive sample augmentation approaches. Experimental results show that the proposed method improves nDCG@5 on ViDoRe v2 to 63.0. Furthermore, we release 3.3 million teacher-annotated samples along with the source code to facilitate future research.

0 citationsRead paper

omni-macos: On-Device Omni-Modal Search on Apple Silicon

Aug 05, 2026

This work addresses the challenge of unifying text, code, documents, images, audio, and video in on-device multimodal search under stringent local compute and memory constraints. It presents the first end-to-end, full-modality multimodal search system deployed entirely on Apple Silicon Macs, integrating encoders, indexing, and storage while ensuring complete data locality. The system innovatively combines multimodal embeddings, quantized indexing, incremental re-encoding, fine-grained GPU acceleration, and memory-aware scheduling, all coordinated through a unified memory allocation mechanism that adheres to user-specified memory budgets. Evaluated across five Mac devices spanning an 8× range in NPU compute and a 32× range in memory capacity, the system consistently achieves efficient background indexing and real-time interactive search performance.

0 citationsRead paper

jina-reranker-v3.5: An Efficient Listwise Reranker with Hybrid Attention and Self-Distillation

Jul 20, 2026

This work addresses the challenge of balancing inference efficiency, cross-domain robustness, and the ability to handle semi-structured data in listwise reranking models. We propose an efficient 0.6B-parameter listwise reranker based on a last-but-not-late (LBNL) interaction architecture, which innovatively integrates sliding-window and global attention mechanisms—comprising three local and two global attention layers—and leverages multi-domain mixed training combined with a three-stage self-distillation strategy. The resulting model achieves 63.20 nDCG@10 on BEIR, matching the performance of 4B-scale models, outperforms prior approaches on MIRACL and RTEB benchmarks, and yields a 9.6-point gain in nDCG@10 on semi-structured retrieval tasks while reducing inference latency by up to 1.56×.

0 citationsRead paper

Test-Time Compute for Dense Retrieval: Agentic Program Generation with Frozen Embedding Models

May 11, 2026

This work addresses the challenge of enhancing retrieval performance of frozen small-scale embedding models without retraining. It proposes a test-time computation approach that leverages agent-based program search to automatically generate and evaluate inference programs for optimizing dense retrieval. The key insight is the discovery of a parameter-agnostic, unified algebraic formulation: interpolating the original query vector with the softmax-weighted centroid of its local top-K retrieved documents. This method consistently improves nDCG@10 across seven embedding model families spanning an order-of-magnitude difference in parameter count and demonstrates broad effectiveness on the full BEIR benchmark suite. To the best of our knowledge, this is the first demonstration that test-time computation can substantially boost the retrieval capability of frozen embedding models.

0 citationsRead paper
Recent publications

Latest Papers

What Transfers from a VLM Teacher? Comparing Supervision Signals for Visual Document Retrieval

Oct 06, 2026

This study addresses the problem of false negative misclassification caused by incomplete annotations in visual document retrieval. We propose a hard negative discrimination and distillation framework leveraging teacher signals from Vision-Language Models (VLMs). By scoring candidate negatives with a VLM, we systematically compare knowledge transfer effects under different supervision signals and optimize the retrieval training strategy through contrastive learning and attention mechanisms. This work is the first to demonstrate that VLM-guided hard negative discrimination significantly outperforms conventional positive sample augmentation approaches. Experimental results show that the proposed method improves nDCG@5 on ViDoRe v2 to 63.0. Furthermore, we release 3.3 million teacher-annotated samples along with the source code to facilitate future research.

0 citationsRead paper

omni-macos: On-Device Omni-Modal Search on Apple Silicon

Aug 05, 2026

This work addresses the challenge of unifying text, code, documents, images, audio, and video in on-device multimodal search under stringent local compute and memory constraints. It presents the first end-to-end, full-modality multimodal search system deployed entirely on Apple Silicon Macs, integrating encoders, indexing, and storage while ensuring complete data locality. The system innovatively combines multimodal embeddings, quantized indexing, incremental re-encoding, fine-grained GPU acceleration, and memory-aware scheduling, all coordinated through a unified memory allocation mechanism that adheres to user-specified memory budgets. Evaluated across five Mac devices spanning an 8× range in NPU compute and a 32× range in memory capacity, the system consistently achieves efficient background indexing and real-time interactive search performance.

0 citationsRead paper

jina-reranker-v3.5: An Efficient Listwise Reranker with Hybrid Attention and Self-Distillation

Jul 20, 2026

This work addresses the challenge of balancing inference efficiency, cross-domain robustness, and the ability to handle semi-structured data in listwise reranking models. We propose an efficient 0.6B-parameter listwise reranker based on a last-but-not-late (LBNL) interaction architecture, which innovatively integrates sliding-window and global attention mechanisms—comprising three local and two global attention layers—and leverages multi-domain mixed training combined with a three-stage self-distillation strategy. The resulting model achieves 63.20 nDCG@10 on BEIR, matching the performance of 4B-scale models, outperforms prior approaches on MIRACL and RTEB benchmarks, and yields a 9.6-point gain in nDCG@10 on semi-structured retrieval tasks while reducing inference latency by up to 1.56×.

0 citationsRead paper

Test-Time Compute for Dense Retrieval: Agentic Program Generation with Frozen Embedding Models

May 11, 2026

This work addresses the challenge of enhancing retrieval performance of frozen small-scale embedding models without retraining. It proposes a test-time computation approach that leverages agent-based program search to automatically generate and evaluate inference programs for optimizing dense retrieval. The key insight is the discovery of a parameter-agnostic, unified algebraic formulation: interpolating the original query vector with the softmax-weighted centroid of its local top-K retrieved documents. This method consistently improves nDCG@10 across seven embedding model families spanning an order-of-magnitude difference in parameter count and demonstrates broad effectiveness on the full BEIR benchmark suite. To the best of our knowledge, this is the first demonstration that test-time computation can substantially boost the retrieval capability of frozen embedding models.

0 citationsRead paper