ScholarPulse 日报 2026-07-07

2026-07-07 学术简报：2 篇。CompactionRL通过上下文压缩技术显著提升长时域智能体LLM的性能。

今日速览

序号	标题	来源	日期	主题	推荐等级
1	CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents	arXiv	2026-07-06	AI-Agent	高
2	PiSAs: Benchmarking Contextual Integrity in Multi-User Agentic Systems	arXiv	2026-07-06	AI-Agent	高

重点论文与技术动态

1. CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents

来源：arXiv
日期：2026-07-06
作者/机构：Yujiang Li, Zhenyu Hou, Yi Jing, Jie Tang, Yuxiao Dong
主题标签：AI-Agent, arXiv
推荐等级：高
分类：cs.LG

一句话结论

CompactionRL通过上下文压缩技术显著提升长时域智能体LLM的性能。

一段话：该方法针对长交互轨迹超出上下文窗口的瓶颈，提出联合优化任务执行与摘要生成的强化学习策略，使LLM能从压缩后的长轨迹中学习，在agentic coding任务上实现一致性能提升。

核心内容

解决长时域LLM因有限上下文窗口导致的交互轨迹过长问题，通过上下文压缩总结先前状态。
采用token级损失归一化和跨轨迹广义优势估计联合优化任务执行与摘要生成。
在SWE-bench Verified和Terminal-Bench 2.0基准上实现Pass@1绝对提升5.5-7.0点，如GLM-4.5-Air达66.8%。

方法与数据

方法：token-level loss normalization and cross-trajectory generalized advantage estimation。
数据：基于GLM-4.5-Air (106B-A30B)和GLM-4.7-Flash (30B-A3B)等开源模型在SWE-bench和Terminal-Bench测试集验证。

价值判断

值得关注：GLM-4.5-Air在SWE-bench Verified达到66.8% Pass@1，较基线提升7.0点，性能显著。
可复用点：方法已部署于GLM-5.2 (750B-A40B)的RL训练管道，可推广至其他开源LLM模型。
局限/待核查：摘要未明确提及局限。

摘要

Long-horizon agentic LLMs are increasingly limited by finite context windows, as extended interaction trajectories can exceed the maximum context length before a task is completed. Context compaction offers a natural solution by summarizing previous interaction states and continuing the rollout under a compressed context, but incorporating compaction into reinforcement learning remains underexplored. We propose CompactionRL, a reinforcement learning strategy to train long-horizon agentic LLMs with context compaction. Our approach jointly optimizes task execution and summary generation with token-level loss normalization and cross-trajectory generalized advantage estimation. This design enables the LLM agents to learn from compacted long-horizon trajectories. We train CompactionRL on top of open models and observe consistent performance gains on agentic coding tasks. CompactionRL enables the open GLM-4.5-Air model (106B-A30B) to achieve Pass@1 scores of 66.8% on SWE-bench Verified and 24.5% on Terminal-Bench 2.0, with absolute gains of 7.0 and 3.1 points, respectively. Built upon GLM-4.7-Flash (30B-A3B), CompactionRL improves Pass@1 by 5.5 and 6.8 points, reaching 56.0% on SWE-bench Verified and 20.2% on Terminal-Bench 2.0, respectively. CompactionRL is thus deployed in the RL pipeline for training the open GLM-5.2 model (750B-A40B).

2. PiSAs: Benchmarking Contextual Integrity in Multi-User Agentic Systems

来源：arXiv
日期：2026-07-06
作者/机构：Shubham Gupta, Nazanin Mohammadi Sepahvand, Abhinav Kumar, Cem Subakan, Spandana Gella, Pierre-André Noël
主题标签：AI-Agent, arXiv
推荐等级：高
分类：cs.MA, cs.CR

一句话结论

多用户LLM代理系统中，LLM判断错误导致隐私泄露风险无法可靠控制，现有模型难以过滤不当内容或限制授权用户访问。

核心内容

PiSAs是首个针对多用户代理系统设计的隐私基准测试工具，通过双重上下文完整性（CI）标注评估信息适用性及合法访问用户权限。
它能直接测量跨用户泄露风险，覆盖系统组件如输出、代理间通信和共享内存等接口。
研究发现，系统设计虽提升CI合规性，但LLM判断错误是主要瓶颈，即使先进模型也失败。

方法与数据

摘要未明确具体方法与数据集。

价值判断

值得关注：多用户代理系统内部信息泄露的新型隐私风险，超越传统单用户场景。
可复用点：PiSAs系统无关性支持跨不同代理拓扑和内存模式评估，便于广泛部署。
局限/待核查：LLM判断错误导致泄露无法控制，需开发新隐私策略以解决核心问题。

摘要

As LLM agents evolve from single-user assistants into shared organizational infrastructure, new privacy risks emerge: inappropriate information may not only be exposed through outputs for external recipients, but also internally across users through inter-agent messages, shared memory and agents. These data spillage risks are not captured by existing privacy benchmarks grounded in contextual integrity (CI) as they focus primarily on either single-user settings or interactions between independently owned agents. We introducePiSAs (Privacy in Shared Agentic systems), a benchmark for assessing unintentional leaks with dual CI annotations: whether an information is appropriate for the task, and which users may legitimately access it. This enables direct measurement of cross-user spillage across agentic system components and interfaces, such as outputs, inter-agent communication, and memory. PiSAsis system-agnostic and supports evaluation across different agent topologies and memory regimes. We find that, although system design improves CI compliance, results are bottlenecked by incorrect LLM judgment calls: even state-of-the-art models fail to reliably filter inappropriate content or restrict transmission to authorized users. Our findings underscore the need for privacy-preserving strategies, beyond those studied in this work.