MS MARCO
PulseAugur coverage of MS MARCO — every cluster mentioning MS MARCO across labs, papers, and developer communities, ranked by signal.
4 day(s) with sentiment data
-
New scaling law for data-constrained dense retrieval models unveiled
Researchers have developed a new scaling law to describe the performance of dense retrieval models under data-constrained fine-tuning conditions. This law, based on the minimum achievable evaluation loss (MAEL), accurat…
-
Glance integrates reranking into generative AI models, reducing overhead
Researchers have developed Glance, a novel approach that integrates reranking capabilities directly into the generative model, eliminating the need for separate reranking models. This method uses a lightweight decision …
-
Glance AI model integrates reranking and generation into single backbone
Researchers have developed Glance, a novel system that integrates reranking and generation into a single AI model. This approach uses a lightweight decision adapter and pointer head to score candidate passages from a sh…
-
New NHNM Method Enhances Instruction-Following Retrieval Models
Researchers have developed a new method called Natural Hard-Negative Mining (NHNM) to improve instruction-following information retrieval (IFIR). This technique repurposes original positive documents as natural instruct…
-
New research explores LLM integration with sparse retrieval systems
Two new research papers explore methods for improving large language model (LLM) integration with sparse retrieval systems. The first paper, "Aligning LLMs with Sparse Retrieval via Adaptive Sparse Supervision," propose…
-
Qwen3-0.6B models achieve SOTA MS MARCO ranking via data curation and Model Soup
Researchers have demonstrated that lightweight Qwen3-0.6B models can achieve competitive passage ranking performance on the MS MARCO benchmark through data-centric strategies rather than architectural changes. Key to th…
-
New AI methods enable scalable embedding model adaptation and Google releases open-source trimodal model
Researchers have developed SYNTRA, a novel method for synthesizing retrieval training data that can adapt embedding models across various domains and tasks with minimal human intervention. This self-improving framework …
-
New method recovers retrieval quality lost to document vector compression
Researchers have developed a method to improve the quality of compressed document vectors in late-interaction models. By learning a global linear correction, the method can recover retrieval quality lost due to aggressi…
-
New methods slash memory use for learned sparse retrieval indexes
Researchers have developed new methods to significantly reduce the memory requirements of learned sparse retrieval indexes without compromising retrieval efficiency. The proposed techniques involve replacing per-block s…
-
New research re-evaluates generative vs. dense retrieval methods · 5 sources tracked
Researchers are exploring new approaches to generative retrieval (GR) and dense retrieval (DR) for tasks like recommendation and question answering. One study found that while GR excels at identifying the best relevant …
-
New method optimizes LLM-based retrieval models for efficiency
Researchers have developed a new method to optimize the sparsity of neural sparse retrieval models, which utilize Large Language Models (LLMs) for semantic term expansion. This approach combines adaptive strategies like…
-
DIY DSSM improves search with translation tables
A developer has created a "poor man's" Deep Structured Semantic Model (DSSM) that uses count-based translation tables to enhance full-text search capabilities. This method enriches the inverted index by associating docu…
-
New Matryoshka Hash Representations improve RAG retrieval efficiency
Researchers have developed Matryoshka Hash Representations (MHR), a novel method for compact semantic retrieval in retrieval-augmented generation (RAG) systems. MHR addresses the challenge of storing large vector indexe…
-
New 'Embedding Surgery' Technique Enhances Dense Retrieval Systems
Researchers have developed a new technique called "embedding surgery" to improve the performance of dense retrieval systems. This method allows for localized, minimal updates to document embeddings at query time, guided…
-
New TRIS defense system combats knowledge poisoning in RAG models
Researchers have developed TRIS, a Tri-Layer Retrieval Integrity Sieve, to combat knowledge poisoning in retrieval-augmented generation (RAG) systems. This middleware defense system sanitizes retrieved evidence by emplo…
-
Supervised fine-tuning impacts LLM instruction sensitivity differently by model scale
A new study published on arXiv investigates how supervised fine-tuning (SFT) affects the instruction sensitivity of large language models. Researchers found that SFT consistently reduces instruction sensitivity in small…
-
New Turkish LLM MoganBert-TR trained with CLM-to-MLM curriculum
Researchers have developed MoganBert-TR, a new Turkish encoder foundation model, and its accompanying embedding model, MoganBert-Embed. Trained from scratch on a filtered Turkish corpus using a novel CLM-to-MLM curricul…
-
New arXiv papers explore privacy, efficiency, and LLM integration in dense retrieval
Four new arXiv papers explore advancements in dense retrieval, a key component for large language models in information retrieval tasks. The first paper introduces a privacy-preserving method using learned deep hashing …
-
Vector database admission control tackles retrieval hubs under workload drift
Researchers have developed a new admission control mechanism for vector databases to mitigate the impact of retrieval hubs, where a single document dominates search results. This system maintains a set of sentinel queri…
-
New research explores static pruning and LLM-based query expansion for retrieval systems
Two new research papers explore methods to improve information retrieval systems. The first paper, "Static Pruning Across Sparse Retrieval Regimes," investigates how static pruning techniques can be applied across diffe…