retrieval

论文提出 SearchWeaver 框架，利用代理的执行轨迹和任务奖励自动联合优化语料表示、可执行工具与使用指令，从而改进 agentic search 的搜索接口设计。使用 GPT-5.4-nano 时，优化后的接口在 BrowseComp-Plus 上将答案准确率从 71.6% 提升至 78.4%，并在四个 BRIGHT 子集上将平均 nDCG@10 从 45.7 提升至 55.0。

KeenableAI（Keenable）凭借自建网页索引与基于 token 相关性打分的分块检索技术，在 Vals Web Search Index 排名升至第一（49.1%），其搜索 API 可通过 MCP 或 API 接入智能体，并能减少约 50% 的输出 token 且不增加延迟。

The author proposes replacing large 'know-it-all' AI models with small models that navigate externally stored knowledge 'maps' built by a retrieval agent, trading memory for storage while keeping results repeatable. The post weighs pros and cons of the concept and invites discussion.

NVIDIA expands its open source Agent Toolkit with the new Agent Knowledge Foundation, adding CUDA-X libraries, Auto-Ontology, Nemotron Retriever and the Kumo family of structured-data models, alongside OpenShell runtime and NemoClaw blueprints, to help enterprises ground AI agents in proprietary company data with governance controls. Early adopters include Palantir, Databricks, Snowflake and Amazon OpenSearch Service.

A practitioner shares hard-won lessons from moving agentic workflows from demos to production, highlighting quiet tool-call failures, goal drift over long chains, missing evals, retrieval issues over model quality, and the value of human approval over autonomy.

OpenAI details how it built a bespoke advertising stack for ChatGPT using a custom post-trained LLM with tool-call architecture for retrieval, ranking, relevance, and auctions, featuring stateful ad sessions and separation from answer generation, having reached $1B annualized revenue with over 100,000 advertisers.

本文评估语言模型智能体的上下文压缩方法在'后续问题'上的表现，发现针对当前问题选择的句子在后续问题上并不优于随机删除，而 LLMingua-2 等词级压缩方法在多个基准上表现更佳。

This paper studies learned post-training quantization calibration for 4-bit text embedders (e5-base-v2, BGE-M3, EmbeddingGemma-300m), finding that calibration outcome ranges from gains to catastrophic collapse depending on the 'dose' of update, and proposes a validation-gated abstention rule that never selects harmful checkpoints while retaining most beneficial ones.

The paper introduces REM (Reusable Experience Memory), a framework that manages experience memory for coding agents through store, recall, and forget decisions inside the think-act loop, outperforming state-of-the-art memory systems on SWE-ContextBench Lite with 68.7% match@1 and 27.6% task resolution at lower cost.

The paper presents ComplianceAgent, the legal compliance sub-agent of the COMPASS multi-agent framework, which uses an LLM to score texts against legal corpora with corrective rewrite loops and human escalation, evaluated on a Canadian and Quebec legal corpus with a hand-annotated gold standard.

LightCRS 提出一个轻量级查询条件化的多域多模态对话推荐框架，仅用5.3M参数模块实现跨域共享的证据融合，在新建的25域 MM-R2A-25 基准上将 HR@1 提升35.4%并将每查询延迟降低266倍。

该论文提出将少量专家(SME)评判转化为密集奖励信号，并与检索指标结合成适用于GRPO的混合奖励，以解决非可验证领域（如金融服务检索）中稀疏奖励导致的梯度消失问题，实验显示Recall@3提升最多6%、MRR@3提升6.7%。

Exa 发布 3.0 版本，称其在全部八项基准测试中排名第一，并新增低价搜索类型 Exa Lite（每千次请求 4 美元），同时将 Dynamic Highlights 正式开放并大幅扩展索引数据。

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.

New Amazon paper introduces FEB-30, a benchmark of 30 scripted fact histories (147 documents) revealing that QA-based agent memory benchmarks can mask broken memory stores, with scores differing 16.8 points on average and verdicts disagreeing on 21% of cases, recommending direct verification of stored facts after each update.

The paper presents MFM, a multimodal filesystem memory architecture that combines episodic, spatial, semantic, procedural, and working memory to help long-horizon embodied agents reuse experience, achieving improved performance on Minecraft and computer-use benchmarks across three base models.

该论文提出了 QEMem 框架，将 LLM 检索的记忆重组为查询特定的证据集合，以弥合“相关性-充分性差距”，在 LoCoMo 基准上以 Claude Opus 4.5 达到 91.43% 的 LLM-judge 准确率，超越 MemR3 和 NEMORI 基线。

论文提出 GRSE（Group-Relative Skill Evolution），通过重复执行并对比同一技能库下的多次运行，将共享团队奖励正确归因到具体的 agent-技能对上，从而改进协作式语言智能体的技能演化，在 HotpotQA 与 LoCoMo 上超越 MASkills 并能从有害技能中恢复。

The paper identifies a Competing Learning phenomenon in hybrid language models where recurrence and attention layers compete to learn the same functions, and proposes a staged training procedure that improves language modeling loss, common-sense reasoning, and retrieval tasks on models up to 1B parameters.

SpecMRL 提出在参数化谱空间中重新组织特征以改进嵌入的前缀读出，在 MTEB 上以更少的嵌入维度超越标准 MRL 及其全维表现。