arXiv:2608. 03397v1 Announce Type: new Abstract: MMLongBench-Doc is a long-document QA benchmark of 1,082 questions over 135 PDFs.
arXiv:2608. 08512v1 Announce Type: new Abstract: Evolving documents, such as laws, tax codes, and software documentation, are amended, replaced, and sometimes reverted over time, so a question has different correct answers at different dates.
arXiv:2609.38021v1 Announce Type: cross Abstract: We evaluate an auditable long-term memory system on LongMemEval-S. Its retrieval chain uses hybrid candidate retrieval, cross-encoder reranking, cove...
AtomCite is an agentic framework that verifies and corrects page‑level citations in multi‑page documents by parsing answers into claims, checking each claim against the cited page image, and applying a deterministic repair policy. The authors introduce DocCite, the first benchmark for this task, built on MP‑DocVQA and DUDE, containing 928 injected instances and 1,909 verified natural errors. Across Gemini, Claude, and GPT models, AtomCite achieves about 93% verification accuracy and improves citation precision from 34% to 87‑90%, while also enhancing hallucination detection in open‑source models.
arXiv:2610.26951v1 Announce Type: new Abstract: Retrieval-augmented generation (RAG) is the dominant paradigm for grounding large language models in external corpora, but its similarity-driven, chunk...
The paper introduces ingest‑time fact compilation, an architecture that preprocesses and compiles corpus data into self‑contained facts with resolved revisions, deletions, and source trust. By storing this compiled state, query‑time models can retrieve answers directly, avoiding costly reconstruction from raw passages. Experiments show that this approach reduces read cost per question by 12.89× and token usage by 21.6× while maintaining accuracy.
KnowForge is a modular pipeline for open‑domain multimodal document question answering that separates retrieval, judging, and reading into distinct agents. It uses a hybrid‑retrieval knowledge‑point index, a Locator agent for searching, a Bridge Judge to enforce a second hop, a Deep Reader that handles text, images, and tables, and a synthesizer with format checking. On the M3DocVQA benchmark it achieves an F1 of 47.5, outperforming prior systems by 11 points, and shows strong gains on MultiHop‑RAG and MMLongBench‑Doc as well.
The paper investigates how conversational Retrieval-Augmented Generation (RAG) benchmarks that rely on relevance judgments to determine context sufficiency can be misleading. By re‑running official baselines on the RETECO track and evaluating with both large‑language‑model judges and a blind audit, the authors show that stronger retrievers produce a higher proportion of contexts lacking gold documents yet still containing the answer, leading to a sharp drop in shortcut precision from 0.68 to 0.10. The study demonstrates that many false misses arise from hard negatives, gold content from other turns, or neighboring chunks, and that using retrieval scores alone as a sufficiency detector can underestimate answerability by up to 0.19 in AUROC.
arXiv:2611. 03196v1 Announce Type: cross Abstract: Dh\={a}r\={a} Bench is a retrieval benchmark for Indian Knowledge Systems (IKS), with 325 questions in 12 categories and 60,121 text chunks from 15 source groups.
SciRerankBench is a benchmark designed to evaluate rerankers in retrieval‑augmented large language models (RAG‑LLMs) for scientific literature question answering. It covers five scientific subjects and introduces three types of question‑context‑answer pairs—Noisy Contexts, Semantically Similar but Logically Irrelevant Contexts, and Counterfactual Contexts—to test noise resilience, relevance disambiguation, and factual consistency. The study evaluates 13 popular rerankers across five LLM families, offering detailed insights into their strengths and limitations.
The paper presents a deployed system for answering questions over normative documents that is aware of document version and scope. It evaluates a hosted retrieval service against a governed system that applies explicit rules for version and scope resolution, finding the governed system achieves a higher score (97.7 vs 88.1). The study includes a public benchmark, evaluation scripts, and reports commercial deployment metrics, such as 1,126 users and 100,000 calls per day by April 2026.
Correctly answering a question grounded in normative documents often depends on information outside any single passage: whether the retrieved document is the version currently in force; whether it app...
The study investigates how document segmentation (chunking) and embedding choices influence Retrieval-Augmented Generation (RAG) performance for Turkish, a morphologically rich language. Using a fully crossed experimental design, the authors compare three chunking strategies (fixed-length, semantic, and layout-aware Docling), five embedding models, and two large language model generators across three documents with different layouts, generating 9,000 graded question-answer evaluations. Key findings include that layout-aware chunking reduces the impact of embedding choice, the top embedding models perform similarly, faster generators are not more accurate, and the optimal configuration varies with content type, achieving a best overall accuracy of 87.0%.