Tag
Tevatron-Elastic presents a unified abstraction for training elastic retrievers and rerankers, enabling a single checkpoint to serve multiple model sizes across depth, token, and width axes. It generalizes prior methods like Matryoshka embeddings and early exit, and introduces Matryoshka LTC for joint token-compression training.
SPIEval is a human-curated benchmark for evaluating LLMs as mobile assistants on scattered personal information tasks, covering 250 tasks across 10 apps. It reveals that even the best model, GPT-5.5 (xhigh), achieves only 57.3% accuracy, with most failures due to inaccurate information localization and insufficient verification.
This paper introduces Factorized Hypothesis Search (FHS), a method for retrieving concepts from large taxonomies when inputs provide indirect contextual evidence, such as table cells or clinical notes. FHS achieves strong results on financial taxonomy tagging and CodiEsp clinical coding, outperforming non-oracle baselines in Recall@1, MRR, and accuracy.
The paper presents TourSynbio-Search, an LLM-driven agent framework for unified protein engineering search across literature and biological databases, powered by the TourSynbio-7B multimodal model with dual PaperSearch and ProteinSearch components.
Introduces LegalPincite, a large-scale legal information retrieval dataset built from CJEU judgments, featuring masked queries, full corpora, and paragraph-level citation annotations to enable multi-level retrieval evaluation.
This paper introduces IFCMemoryBench, a human-validated benchmark for evaluating long-term memory in LLM-based agents for BIM information retrieval. It shows that current memory systems achieve only 32.4% answer accuracy, revealing a domain-transfer gap in agent memory.
GuidedRAG is a research paper proposing a novel RAG extension that adds a semantics-based selection stage before retrieval, improving retrieval relevance and reducing overhead across diverse RAG variants.
This paper presents a controlled scaling study comparing lexical, dense, graph-based, and agentic RAG paradigms across corpus sizes from 1,000 to 512,000 documents, finding that BM25 provides the best accuracy-cost tradeoff, while graph-based RAG faces high construction costs that limit scalability.
This paper investigates how the size of text chunks (sentences, paragraphs, chapters) in Retrieval-Augmented Generation systems affects generation quality, retrieval precision, and computational efficiency, using identical textbooks segmented at different granularities.
This paper systematically investigates Semantic IDs (SIDs) in generative recommendation, finding that while SIDs preserve coarse item organization, they lose fine local structure from the encoder. The authors propose Item-Supported Decoding (ISD), a lightweight inference-time method that improves NDCG@10 by up to 31.2% without additional parameters or retraining.
AWS highlights that 58% of employees lose 2 hours daily searching for information, and promotes Amazon Quick as a solution to reclaim lost time.
This paper proposes improvements to HyperGraphRAG by using self-consistency prompting for better fact extraction and Personalized PageRank for enhanced chunk retrieval.
This paper presents adaptive pipelines for legal retrieval, entailment, and judgment prediction tasks in the COLIEE 2026 competition, using multi-stage retrieval, reranking, and LLM-based reasoning.
AutoIndex is a framework that learns executable representation programs to transform documents before indexing, improving retrieval quality over static BM25 baselines by average 8.4% Recall@100 and 8.3% nDCG@10.
This paper quantifies and expands the theoretical capacity of late-interaction retrieval models, showing that MaxSim can replicate inner products between non-negative vectors and proposing Signed MaxSim for arbitrary real-valued vectors, revealing a representation gap between inner product and late-interaction models.
The recommended podwise-cli is a CLI tool that converts any podcast content into structured summaries, key insights, and mind maps, making it convenient for AI agents and workflows.
The paper presents Libra, a self-evolving framework that introduces mutable catalogs into repositories to improve code localization for LLM agents, achieving logarithmic improvements and zero-shot transfer across different models and problem sets.
This paper identifies a vocabulary gap as the root cause why advanced encoders like ModernBERT underperform in learned sparse retrieval, and proposes Vocabulary Transfer (VT), a model-agnostic framework that migrates encoders to sparse-friendly vocabularies, achieving state-of-the-art on the BEIR benchmark.
This paper presents a Bayesian uncertainty propagation framework for Agentic RAG systems, evaluating it on multi-hop QA benchmarks with GPT models, showing promise for monitoring reliability in industrial decision support.
This paper introduces PaperPilot, a multi-turn literature search agent that constructs executable DAGs of search operators and refines workflows through user feedback, achieving significant improvements in retrieval metrics over baseline models.