information-retrieval

Meta's DecisionRank paper introduces item-space on-policy distillation to transfer a generative LLM teacher's ranking knowledge into a single-token decision-model reranker, achieving state-of-the-art results on TREC DL, BEIR, and BRIGHT while running up to 24.5x faster than its generative counterpart and surpassing its 27B teacher at just 1.7B parameters.

This paper shows that using incomplete relevance judgments as a shortcut for scoring abstention in conversational RAG benchmarks systematically distorts evaluation: stronger retrievers yield more 'false miss' contexts that still contain answers, badly deflating answerability AUROC. The authors release code, runs, and labels and recommend scoring refusals against content-based labels instead of gold-document presence.

IBM researchers introduced STAIR, a retriever that leverages document tables of contents for structured retrieval, achieving 82.6% top-1 section accuracy on the SearchTome benchmark and 65× fewer invalid section references than a fine-tuned DSI baseline.

介绍康奈尔大学与 ASAPP 的 DecisionRAG 论文：用决策模型（如 Jev 1.13）在文档树上导航取代向量相似度检索，实现 retriever-free RAG，在 Natural Questions 上 top-1 命中率达 65.4%，优于 embedding 检索加 8B reranker 的 48.0% 和 GPT-5 reranking 的 60.8%。

DecisionRank 使用 item-space on-policy distillation 将生成式教师模型的排序知识迁移到高效决策模型，在 TREC DL、BEIR、BRIGHT 等检索基准上达到最先进性能，同时推理速度显著提升（最高快 24.5 倍），1.7B 参数即可超越 27B 生成式教师。

Glance integrates reranking and generation into a single 9B backbone using lightweight decision adapters and pointer heads that score passages from a shared prefill, reducing memory overhead while beating bge-reranker-v2-m3 and cross-encoders on MS MARCO benchmarks and HotpotQA answer F1.

Capital One's AI Foundations team introduces UCVE, a domain-agnostic framework that models compliance requirement spaces as a Contextual Requirement Graph (CRG) and performs retrieval via spectral reranking over a reinforcement-learning-weighted Laplacian. Evaluated across three compliance domains, UCVE achieves a 4.3% matched-K recall lift over flat cosine retrieval, outperforming strong retrieval and graph-based baselines.

该论文提出 Parametric Session State (PSS)，一种将长对话交互编译为固定形状的 LoRA 会话记忆层，使 RAG 系统在无需重放历史的情况下恢复未明确表述的查询中的检索意图，并在受控与公开评测中验证其跨模型与检索后端的有效性。

研究比较了 System One（决策）模型与标准 reranker 微调在检索重排任务上的表现，发现决策微调在所有 9B 以下 Qwen 主干上均不如直接进行 reranker 微调，nDCG@10 差距达 5 到 29 分，结论是如果需要重排器，就应该直接微调重排器。

该论文研究了古典阿拉伯语诗歌档案的检索问题，基准测试了九个稠密编码器与BM25，发现查询所保留的原文词句多少决定了哪种检索器占优：在记忆模糊查询下BM25胜过包括8B模型在内的所有稠密编码器，而在抽象查询下表现大幅下滑，说明单一评测协议或查询类型会给出严重失真的结论。

论文提出 Compositional Concept Unlearning (C2U) 方法，在固定编码器与文档索引的 RAG 系统中通过查询侧表示编辑来抑制检索到不需要的概念，跨五个数据集和四个编码器验证了其在不损害控制查询效用的情况下降低目标概念的检索暴露。

This paper compares 11 retrieval configurations for long-document RAG—including lexical, dense, hybrid reranking, and full-document scanning—evaluating evidence coverage, answer quality, and cost on QASPER and textbook questions, finding that hybrid approaches often outperform full scans.

论文提出联合约束聚合（JCA）方法，通过将问题要求表示为共享约束并逐一核对候选答案，缓解检索偏向导致的限定词敏感问答失败，在三种评测设置和多个指令微调模型上提升准确率。

The paper argues that multimodal reranking can be framed as a single-pass System One choice question for a general vision-language model, showing zero-shot Qwen3.8-27B matching or beating dedicated supervised rerankers on several retrieval benchmarks.

The post explains why RAG needs both embedding-based retrieval and reranking: dual-encoder models are fast and allow offline precomputation of document vectors, making them suitable for large-scale retrieval but weaker at capturing fine details; cross-encoders score each pair more accurately but are computationally expensive, so recall-then-rerank is a necessary trade-off.

SAG2 提出一种无需嵌入的多跳检索架构，仅使用 SQLite、BM25 和 JEV 上下文证据决策，在 MuSiQue、2WikiMultiHopQA 和 HotpotQA 上取得最高 Recall@5/Recall@10，索引速度快且无需模型调用。

一款 GitHub 上的开源 AI 新闻雷达项目，聚合 X、Telegram、Reddit、GitHub、Hacker News、RSS 等多源信息，利用 AI 自动过滤、评分、去重，并生成中英文每日简报。

LongR 是一个统一框架，通过将“Think-and-Read”动态机制与基于相对信息增益的上下文密度奖励相结合，提升 LLM 的长上下文推理能力，在 LongBench v2 上取得超过 9% 的提升，并在 RULER 和 InfiniteBench 上保持一致改进。

一篇实验性博客文章提出「Bag of Decisions」重排方法：让 LLM 为每个查询生成一组是/否判准问题，再用 Jev 计算各问题为真的概率并求和作为相关性分数，在 Wayfair WANDS 和 Amazon ESCI 数据集上均优于 BM25 和单一问题重排。

Maxime Rivest 介绍了一个可直接供 AI agent 调用的语义搜索 API，覆盖 2008 年至今的科学文献，只需发出 GET 请求即可获得检索结果。目前仍是 demo 版本，呼吁开发者试用并反馈改进建议。