今日速览
| 序号 | 标题 | 来源 | 日期 | 主题 | 推荐等级 |
|---|---|---|---|---|---|
| 1 | Learning from Failures: Heterogeneous Graph Memory for Small Language Model Tool-Using Agents | arXiv | 2026-09-23 | AI-Agent | 高 |
| 2 | The KV Cache Working Set: Online Capacity Planning for LLM Inference Systems | arXiv | 2026-09-23 | AI-Agent | 高 |
重点论文与技术动态
1. Learning from Failures: Heterogeneous Graph Memory for Small Language Model Tool-Using Agents
- 来源：arXiv
- 日期：2026-09-23
- 作者/机构：Jiaxing Li, Lei Song, Rui Dong, Youyong Kong
- 主题标签：
AI-Agent,arXiv - 推荐等级：高
- 分类：cs.LG
一句话结论
FRESH 将工具智能体的成功与失败组织为异构图记忆，帮助冻结小/中语言模型在长程有状态任务中复用策略并减少重复失败。
核心内容
- 小/中语言模型成本低，但易漏观察、过早写入、重复失败调用、违反前置条件。
- 微调需大量数据算力，扁平记忆可能检索失败动作，却缺因果上下文与安全条件。
- FRESH 显式建模任务、动作、错误、修复与执行条件依赖，将经验转为外部结构化记忆。
方法与数据
- 方法：提出 FRESH，用于冻结语言模型的工具交互。
- 数据：在 τ-Bench 与 AppWorld 上用多个开源模型，比较无记忆和记忆基线。
价值判断
- 值得关注：面向本地与大规模部署，提升任务成功率与工具可靠性，降低微调依赖。
- 可复用点：可借鉴成功/失败经验结构化与错误—修复—条件依赖的图记忆。
- 局限/待核查：摘要未明确图构建、检索机制、模型规模、成本及不可逆后果量化。
摘要
Small and medium-sized language models offer cost-effective executors for tool-using agents, making them attractive for local and large-scale deployment. However, in long-horizon and stateful environments, they often make structural errors such as missing required observations, performing premature writes, repeating failed calls, and violating action preconditions. These errors can lead to incorrect state updates, policy violations, and costly or irreversible consequences, making reliable tool execution a critical deployment challenge. Existing fine-tuning approaches require substantial data and computation, while flat memory may retrieve failed actions without preserving their causal context or safety conditions. In this paper, we propose FRESH, a Failure-aware Retrieval framework over Experience-Structured Heterogeneous graphs, which transforms historical successes and failures into structured external experience for tool-using agents. By explicitly modeling the dependencies among tasks, actions, errors, repairs, and execution conditions, FRESH helps frozen language models reuse reliable strategies, avoid recurring failures, and make safer decisions in stateful tool interactions. Experiments on $τ$-Bench and AppWorld with multiple open-source models show that FRESH consistently improves task success and tool-use reliability over no-memory agents and representative memory-based baselines.2. The KV Cache Working Set: Online Capacity Planning for LLM Inference Systems
- 来源：arXiv
- 日期：2026-09-23
- 作者/机构：Luchang Li, Shuaishuai Wang, Zhao Ruan, Dongfang Li, Bozhao Gong
- 主题标签：
AI-Agent,arXiv - 推荐等级：高
- 分类：cs.DC
一句话结论
KVSET 是面向 LLM 推理的在线分析器，可估计达到目标 KV 缓存命中率所需的最小容量，辅助容量规划。
核心内容
- 前缀缓存复用历史前缀的 KV 状态，避免重复 prefill，适合上下文增长的 agentic 负载。
- KV 缓存工作集是达到目标命中率所需的最小容量；过大成本高，不足会降低命中率。
- KVSET 用 Mattson 栈算法估计不同容量下的命中率，并确定最小容量。
方法与数据
- 对每个 KV 页计算 LRU 栈距离，并与候选容量页号比较，避免逐容量模拟，降低开销。
- 用生产 LLM 负载 trace 验证，结果接近真实缓存部署；开源支持在线处理与离线回放。
价值判断
- 值得关注：将容量规划从离线试算转为在线分析，适合前缀复用频繁变化的 LLM 服务。
- 可复用点：LRU 栈距离与候选容量比较的命中率估计，以及目标命中率最小容量判定。
- 局限/待核查：摘要未明确命中率目标、页大小、模型规模、硬件环境与误差范围。