Tom 文献雷达 · Agent RAG Long-Context · 2026-10-01

本期候选(8 条)

# 来源 标题 标签 信号
1 arXiv Towards Semi-Automatically Comparing Keyword-Based and Semantic Search Accuracy rag, benchmark 新方法论文
2 HF Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression rag, long-context, memory 97票
3 HF Org-Agent: Beyond Personal Assistants Towards Organizational Agents agent, memory org-level agent 新范式
4 HF Scaling Properties of Same-Family On-Policy Distillation systems 213票,OPD scaling
5 HF Fractional State Space Transition for Long Sequence Modeling long-context, memory power-law long memory SSM
6 HF Can Agents Design Libraries for Agents? agent, benchmark agent library design benchmark
7 HF What Makes Recurrence Effective in Looped Language Models? benchmark, systems LoopLM scaling 分析
8 HF PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents multimodal 多模态 critic search

高价值条目(3 条)

[1] Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression https://arxiv.org/abs/2609.36322 Chunked KV-cache 压缩引入"相位敏感"问题:同一信息在窗口边界的不同相对位置,检索准确率差异显著(高达数十个百分点)。对 RAG 长上下文系统的 KV-cache 设计有直接警示意义。

[2] Org-Agent: Beyond Personal Assistants Towards Organizational Agents https://arxiv.org/abs/2609.34392 超越个人助手边界:org-level agent 需处理跨用户交互/决策、跨用户记忆与知识使用,受用户身份、权限、信息归属与时效性等约束。多 agent 协作系统的工程参考。

[3] Fractional State Space Transition for Long Sequence Modeling (FRAC) https://arxiv.org/abs/2609.36314 用分数阶动力学替代 ODE 指数衰减,实现 power-law 长记忆 SSM,在长时间范围建模上有潜力。对长上下文架构选型有参考价值。

行业动态(Substack/文章,1 条)

RAG in 2026: long context did not kill retrieval, it raised the bar 来源:Slash Digital Lab,2026年9月 https://slash-digital.io/en/insights/rag-2026 要点:1M token 窗口下 RAG 并未过时,混合架构(hybrid search + reranker + agentic search + 大上下文)是 2026 年主流;低于约 20 万 token 稳定内容,cached prompt 已是有效替代方案。


Tom Radar · Agent RAG Long-Context · 2026-10-01 · 8条候选 · 3条高价值 · 1条Substack