文章背景与核心概要
当大语言模型（LLM）系统面临上下文窗口的限制时，通常会通过压缩先前的内容来维持当前任务的推进。然而，这一过程经常会遗漏“会话约束（Session Constraints, SCs）”——即用户在整个会话期间用于规范LLM行为的侧面指令（例如：“在我确认之前不要删除任何邮件”）。本文作者引入了名为 COMPINT 的评估框架，旨在测试压缩器在多轮对话、智能体轨迹和长视野研究这三种长文本场景下的表现。
研究发现，当前的压缩器平均只能保留17%注入的SC，其表现甚至往往比不进行任何压缩直接运行任务还要糟糕。这种约束丢失是一个系统性缺陷，其保留率随压缩器类型、提示词设计、上下文长度、SC表述方式以及注入位置的不同而剧烈波动。为了解决这一问题，作者提出了一种具备SC感知能力的提取器，作为即插即用的旁挂模块（sidecar module）。该模块在无需修改底层压缩器或LLM的情况下，在所有三个评估场景中实现了超过90%的保留率。
Lost in Compaction: Evaluating Side-Constraint Loss under Context Compaction
arXiv: 2608.11242 [cs.CL]
Submitted on: 31 July 2026
Authors: Zhiqi Wang, Yichi Zhang, Dongwon Lee, Yuchen Yang
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Links: View PDF | HTML Version | GitHub Repository
📌 Summary
当大语言模型（LLM）系统面临上下文窗口限制时，它们通常会压缩先前的上下文以继续进行中的任务。然而，这一过程经常会丢弃会话约束（Session Constraints, SCs）——这是用户发布的、旨在规范LLM在整个会话期间行为的侧面约束（例如：“在我确认之前不要删除任何邮件”）。
为了衡量这一现象，作者引入了 COMPINT，这是一个旨在跨三种长上下文场景测试压缩器的评估框架： 1. 多轮聊天 2. 智能体轨迹 3. 长视野研究
核心发现
- 保留率差： 当前的压缩器平均仅保留 17%注入的SC，其表现通常比在没有任何压缩的情况下运行任务还要差。
- 系统性丢失： 保留率随压缩器、提示词设计、上下文长度、SC措辞和注入位置的不同而急剧变化，证明这种丢失是一个系统性缺陷，而非孤立事件。
- 提出的解决方案： 作者引入了一个 SC感知提取器，它作为即插即用的旁挂模块运行。在不修改基础压缩器或LLM的情况下，该模块在所有三个评估场景中均实现了超过 90%的保留率。
When Large Language Model (LLM) systems face context window limitations, they often compact prior context to keep ongoing tasks moving forward. However, this process frequently drops Session Constraints (SCs)—user-issued side-constraints meant to govern the LLM's behavior throughout a session (e.g., "do not delete any emails until I confirm").
To measure this phenomenon, the authors introduce COMPINT, an evaluation framework designed to test compactors across three long-context scenarios: 1. Multi-turn chat 2. Agentic trajectory 3. Long-horizon research
Key Findings
- Poor Retention: Current compactors retain only 17% of injected SCs on average, frequently performing worse than running tasks without any compaction.
- Systemic Loss: Retention varies sharply depending on the compactor, prompt design, context length, SC phrasing, and injection location, proving that the loss is a systematic flaw rather than an isolated incident.
- Proposed Solution: The authors introduce an SC-aware extractor that acts as a plug-and-play sidecar module. Without modifying the base compactor or LLM, this module achieves over 90% retention across all three evaluated scenarios.
📄 Abstract
当上下文窗口承受压力时，LLM系统会压缩先前的上下文以继续进行中的任务。我们识别出一类由用户发出的指令，即会话约束（Session Constraints, SCs），例如“在我确认之前不要删除任何邮件”，这些指令旨在约束LLM在剩余会话中的行为，但在压缩过程中被静默丢弃。为了量化这种丢失，我们引入了COMPINT，这是一个评估套件，用于在三种长上下文场景中评估压缩器：多轮聊天、智能体轨迹和长视野研究。
当前的压缩器平均仅保留17%的注入SC，且大多数压缩器的表现比不进行压缩运行同一任务还要差。保留率随压缩器、提示词、上下文长度、SC措辞和注入位置的不同而急剧变化，这表明丢失是系统性的，而不仅仅与某个单一设置相关。我们提出了一种SC感知提取器，它与压缩器并行作为即插即用模块运行，在不修改压缩器或LLM的情况下，在所有三个场景中实现了超过90%的保留率。
When the context window is under pressure, LLM systems compact prior context to continue ongoing tasks. We identify a class of user-issued instructions, Session Constraints (SCs), such as "do not delete any emails until I confirm," that are meant to constrain LLM's behavior for the remainder of a session but are silently dropped during compaction. To quantify this loss, we introduce COMPINT, an evaluation suite that evaluates compactors across three long-context scenarios: multi-turn chat, agentic trajectory, and long-horizon research.
Current compactors retain only 17% of injected SCs on average, and most perform worse than running the same task without compaction. Retention varies sharply with compactor, prompt, context length, SC phrasing, and injection location, showing that the loss is systematic rather than tied to any single setting. We propose an SC-aware extractor that runs alongside the compactor as a plug-and-play module, achieving over 90% retention across all three scenarios without modifying the compactor or LLM.
🔗 Additional Resources & Citations
- DOI: 10.48550/arXiv.2608.11242
- External Tools: NASA ADS | Google Scholar | Semantic Scholar
- DOI: 10.48550/arXiv.2608.11242
- External Tools: NASA ADS | Google Scholar | Semantic Scholar
