01 · Q2D-Web: A Large-Scale Benchmark for Retrieval in Agentic RAG Systems

Q2D-Web：面向智能体检索增强生成（RAG）系统的大规模检索基准

Research record	Details
Authors	Maximilian Schall, Sedigheh Eslami, Markus Krimmel, Antoine Chaffin, Louis Milliken, Bo Wang, Denis Bykov
Published	2026-09-08
Sources	arxiv
Focus	agentic RAG benchmark, large-scale retrieval, agentic query reformulation, multilingual evaluation 智能体RAG基准, 大规模检索, 智能体查询重写, 多语言评估

English

Q2D-Web provides a rare, large-scale agentic-query corpus and multi-signal labeling methodology directly relevant to evaluating first-stage retrievers in agentic RAG pipelines; both the dataset design and subcorpus-sampling techniques are practically applicable to the thesis’s experimental setup and merit careful study.

中文

Q2D-Web 提供了罕见的大规模智能体查询语料与多信号标注方法，直接关联于智能体 RAG 流水线中一阶段检索的评估；其数据设计和子语料抽样技术对论文的实验设置具有实用价值，应予以深入研读。

Research synopsis

English

This paper introduces Q2D-Web, a large-scale benchmark tailored to first-stage retrieval in agentic RAG pipelines. Q2D-Web pairs a 190M-document web corpus with 70k machine-reformulated agentic queries in ten languages and supplies three fixed relevance-judgment sets: agent citations, production rankings, and a combined set augmented by LLM judgments to reduce false negatives. The authors benchmark 13 retrievers (lexical, dense, late-interaction) and observe that relative model ordering is largely insensitive to the choice of judgment set but varies across domains, languages, and query types. They also propose subcorpus sampling via reciprocal rank fusion to approximate full-corpus evaluation while preserving model ranking.

中文

本文提出 Q2D-Web，一个针对智能体检索增强生成（RAG）流水线一阶段检索的大规模基准。Q2D-Web 将一个包含 1.9 亿文档的网络语料与 7 万条十种语言的机器重写智能体查询配对，并提供三套固定的相关性判断：智能体引用、生产排名，以及通过 LLM 判断增强以减少漏标的合并集合。作者基准测试了 13 种检索器（词法、稠密、后交互），发现模型相对排序对判断集选择总体不敏感，但在专题、语言和查询类型间有显著差异。他们还提出基于互惠秩融合的子语料抽样方法，可在保留模型排序的同时近似全语料评估。

Thesis relevance

English

Overlap: Q2D-Web directly addresses first-stage retrieval in agentic RAG scenarios and provides a large, realistic distribution of machine-reformulated queries and multi-signal relevance labels that can inform evaluation design in the thesis. Differences/limitations: the benchmark targets large web-scale corpora and retriever ranking; it does not evaluate structured, persistent graph memory, multi-hop graph traversal, or entity-resolution effects central to the thesis. Complementarity: the dataset and judgment methodologies (agent citations, production rankings, LLM-augmented pooled labels) and subcorpus sampling technique could be reused to stress-test the thesis’s retrieval component or to create comparable large-scale retrieval baselines.

中文

重合点：Q2D-Web 针对智能体 RAG 场景中的一阶段检索，提供了大规模、真实的机器重写查询分布以及多信号相关性标注，可为论文的评估设计提供借鉴。差异/局限：该基准聚焦于网页规模语料和检索器排序，不评估论文核心的结构化持久图记忆、多跳图遍历或实体解析影响。互补性：其标注方法（智能体引用、生产排名、LLM 增强的 pooled 标签）及子语料抽样技术可用于检验论文中检索子系统的稳健性或构建可比较的大规模检索基线。

Writing and related work

English

Position this work among related evaluation resources: it emphasizes the importance of using machine-reformulated (agentic) queries and multiple relevance signals when evaluating first-stage retrievers. The subcorpus-sampling pragmatic lesson (retain ~1/3 via RRF) is a useful engineering note for large-scale experiments.

中文

将此工作归入相关评测资源时强调：评估一阶段检索器时应使用机器重写（智能体）查询并采纳多重相关性信号。关于工程实践的要点是子语料抽样（通过 RRF 保留约 1/3）可显著加速大规模实验。

Method and evaluation

English

For the thesis, adopt the paper’s evaluation practices: use agentic query distributions rather than only human queries, include multiple relevance signals (agent citations, production rankings, and pooled LLM judgments) to reduce false negatives, and benchmark across lexical/dense/late-interaction retrievers. Use reciprocal-rank-fusion-based subcorpus selection to speed repeated experiments while checking that model rankings remain stable under your task and judgment set.

中文

对于论文，可借鉴其评估实践：采用智能体查询分布而非仅有人类查询，结合多重相关性信号（智能体引用、生产排名及 pooled LLM 判断）以降低漏标，并在词法/稠密/后交互检索器间进行比较。使用基于互惠秩融合的子语料选择来加速重复实验，同时验证在你的任务和判断集下模型排序是否稳定。

Future directions

English

Extend this benchmark paradigm to cover multi-hop questions with explicit bridge-entity annotations and to include retrieval workloads tailored to local or privacy-preserving models. Another direction is to add evaluation suites that measure the impact of entity-resolution and structured graph memory on downstream multi-hop QA.

中文

可将该基准范式扩展为包含显式桥接实体注释的多跳问题，并纳入针对本地或隐私保护模型的检索任务。另一方向是增加评测套件，用以衡量实体解析和结构化图记忆对下游多跳问答的影响。

---

02 · Personalizing LLM Agent Memory Using Biometrics

基于生物特征的 LLM 智能体记忆个性化

Research record	Details
Authors	Yanhong Qian, Qingguo Meng, Shihao Ding, Xingbo Dong, Zhe Jin, Hanrui Wang, Isao Echizen
Published	2026-09-08
Sources	arxiv
Focus	biometric-aware memory, personalized retrieval, multi-user agent memory, memory retrieval gating 生物特征感知记忆, 个性化检索, 多用户 智能体记忆, 检索候选池预筛选

English

Relevant for access-control and personalization aspects of agent memory but not central to the thesis’s focus on query-aware structured graph memory, entity resolution, and multi-hop retrieval; read for ideas on user conditioning and privacy controls rather than core methods.

中文

该文对智能体记忆的访问控制与个性化有参考价值，但与论文关注的查询感知图记忆、实体消歧和多跳检索并非核心重合；建议阅读以获取用户条件化与隐私控制的思路，而非核心方法细节。

Research synopsis

English

The paper addresses personalization of LLM agent memory in multi-user settings by introducing Bio-Memory, a biometric-aware memory architecture. Built on A-Mem, Bio-Memory augments each atomic memory note with a biometric embedding and uses biometric matching to form a retrieval candidate pool which is then ranked by semantic similarity. Evaluation on LoCoMo in a 10-user shared-agent setting across seven face benchmarks and ten palmprint protocols shows consistent separation between owner and non-owner queries; reported gaps include up to 27.29% F1 and 21.15% BLEU-1 (face) and 25.75% F1 and 19.22% BLEU-1 (palmprint). The contribution is a practical workflow for conditioning agent memory retrieval on biometric matching to improve owner-specific recall.

中文

本文研究多用户场景下对 LLM 智能体记忆的个性化，提出了 Bio-Memory——一种生物特征感知的记忆架构。在 A-Mem 基础上，Bio-Memory 为每条原子记忆附加生物特征嵌入，先基于生物特征匹配构建检索候选池，再对候选项按语义相似度排序。在 LoCoMo 的 10 用户共享智能体设置、7 个人脸基准和 10 个掌纹协议上评估，作者报告了稳定的所有者/非所有者区分效果；示例差距包括人脸上最高 27.29% F1 和 21.15% BLEU-1，掌纹上分别为 25.75% 和 19.22%。该工作展示了将生物特征作为智能体记忆检索控制信号的可行性。

Thesis relevance

English

Overlap: both works concern mechanisms to gate or condition agent memory retrieval in multi-user contexts and aim to improve correct reuse of past memory items. Difference: this paper focuses on biometric gating of flat atomic memories rather than persistent structured graph memory or multi-hop reasoning over an assembled knowledge graph (KG). Limitations relative to the thesis: it does not address graph construction, entity/relation extraction, entity resolution strategies, or multi-hop path traversal and verifiability. Complementarity: biometric-based candidate filtering could be combined with query-aware graph memory to restrict node/edge access by user identity and to study privacy/ownership effects on multi-hop retrieval and graph coherence.

中文

重合点：两项工作都关注在多用户场景下对智能体记忆检索的条件化或门控，以提升对过往记忆项的正确重用。差异：该论文侧重于对扁平原子记忆的生物特征门控，而非构建持久的结构化图记忆或在组装的知识图谱（KG）上进行多跳推理。相对于论文课题的局限性：它未讨论图构建、实体/关系抽取、实体消歧策略或多跳路径遍历与可证实性。互补性：可将生物特征候选过滤与查询感知图记忆结合，用以基于用户身份限制节点/边访问，并研究这对多跳检索与图一致性的影响。

Writing and related work

English

Position this paper in related work on access control and personalization for agent memory rather than in structured KG construction. Cite it when discussing user-aware retrieval filters, privacy-preserving memory access, or multi-user shared-agent settings.

中文

将该论文归入关于智能体记忆的访问控制与个性化的相关工作，而非结构化知识图谱构建。讨论用户感知检索过滤、隐私保护的记忆访问或多用户共享智能体场景时可引用。

Method and evaluation

English

Consider integrating biometric embeddings as ownership attributes on graph nodes or edges so that graph memory retrieval first filters candidates by biometric match before performing multi-hop traversal. Experimentally, simulate a multi-user MultiHop-RAG setting by assigning node ownership and measure effects on bridge-entity identification, graph fragmentation, retrieval accuracy, and end-to-end latency. Also evaluate privacy and failure modes (false accepts/rejects) and measure how biometric gating interacts with entity-resolution strategies (e.g., embedding-similarity vs LLM-as-judge).

中文

可将生物特征嵌入作为图节点或边的所有权属性，将图记忆检索先按生物特征匹配筛选候选，再执行多跳遍历。在实验上，可通过为节点分配所有者来模拟多用户 MultiHop-RAG 设置，衡量其对桥接实体识别、图碎片化、检索准确性和端到端延迟的影响。还应评估隐私与失效模式（误接受/误拒），以及生物特征门控与实体消歧策略（如嵌入相似度 vs LLM-as-judge）的交互效果。

Future directions

English

Explore privacy-preserving biometric techniques (secure enclaves, template protection) and non-biometric ownership signals for shared-agent memory. Extend Bio-Memory concepts to structured graph memory to study how owner-conditioned filtering affects multi-hop reasoning, verifiability, and long-term graph coherence.

中文

探索隐私保护的生物特征技术（安全环境、模板保护）以及用于共享智能体记忆的非生物特征所有权信号。将 Bio-Memory 思路扩展到结构化图记忆，研究基于所有者的筛选如何影响多跳推理、可证实性与长期图一致性。

---