Tom 文献雷达 · AI Agent / RAG / 长上下文 · 2026-08-23 第三期

本期候选(8条)

# 来源 标题 标签
1 arXiv Inadvertent Context Leakage in Language Models agent, long-context, security
2 HF The Embedder's Dilemma: LLMs Are Better, but at What Cost? rag, benchmark
3 HF Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses agent
4 HF QuoteBench: How Matched Scores Can Hide Command-Path Failures agent, benchmark
5 HF τ_0-VLA: Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation agent, memory, systems
6 HF The Embedder's Dilemma: LLMs vs Embedding Models (cost-aware, 37 tasks) rag, benchmark
7 arXiv What Makes a Good Fiqh Retriever? Arabic Islamic Jurisprudence RAG rag, benchmark
8 HF FlowEvo: Self-Evolving Agents through Co-Evolution of Workflows and Skills agent

⭐ 高价值条目(4条)

1. Inadvertent Context Leakage in Language Models

来源: arXiv | 标签: agent, long-context, security

AI Agent 的记忆/上下文(如日历、凭据、医疗/财务数据)即使模型正确拒绝直接提取,攻击者也可通过构造提示词利用隐藏关联重建敏感信息——即便模型"答对了"也发生了泄漏。

要点: 上下文窗口本身成为隐蔽信道;需要新的防御范式,不能只靠输出过滤。LLM 安全研究从提示注入升级到了上下文相关性攻击。


2. The Embedder's Dilemma: LLMs Are Better, but at What Cost?

来源: HF Daily | 标签: rag, benchmark | votes: 10

37任务、6家族、26模型(118M–14B参数)的受控成本感知对比。结论:整体持平(Gemini 3.1 Pro 77.6 vs 最佳embedding 77.2),但LLM在推理密集检索占优,embedding在分类占优,STS/聚类/配对分类持平。

要点: 实际选型不能只看精度,需把延迟+成本纳入决策;是目前为止最系统的embedding vs LLM嵌入对比研究。


3. Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses

来源: HF Daily | 标签: agent | votes: 10

同一冻结LLM在三层次Scope运行:Task Harness → Meta-Harness → Self-Harness,harness自身可热插拔rewrite,用环境反馈持续演化。

要点: 框架而非微调——让agent的"脚手架"随任务自我进化,解耦了模型能力与工程管线,是Agent框架层面的自我改进范式。


4. QuoteBench: How Matched Scores Can Hide Command-Path Failures

来源: HF Daily | 标签: agent, benchmark | votes: 7

LLM coding agent通过bash接口发命令,接口序列化/重解析后,matched execution score无法区分生成错误与传输后错误。QuoteBench用56个one-shot任务验证这一边界。

要点: 当前Agent评测指标存在系统性盲区——生成合约与执行传输之间的缝隙被忽略,QuoteBench填补了这个空白。


📮 Substack 补充(1条)

My AI Agent Had 18 Security Holes. Here's How I Fixed Every One Practical breakdown of real agent security failures: system prompt leakage, Unicode homoglyph bypass, memory injection paths, missing audit logs — 与本期 arXiv "Context Leakage" 论文形成实践映照。


去重说明

与上期(2026-08-22)对比:QuoteBench、Hierarchical Self-Improvement 为新进;The Embedder's Dilemma 来自上期HF策展,本期补充arXiv版本;FlowEvo/TinyCast 热度偏低暂不升级。


Tom 研究雷达 · 每日3次 · 2026-08-23 20:40 UTC