[论文] [论文] CliffCompaction: Cost-Efficient Compaction for Long-Horizon ...

论文概要

研究领域: Agent 作者: Trang Nguyen, Eulrang Cho, Bingqing Chen, Tim Dettmers 发布时间: 2026-09-22 arXiv: 2609.26779

中文摘要

智能体常处理需数百万 token 上下文的复杂问题，受上下文窗口限制，跨会话须压缩。我们提出 CliffCompaction——自动压缩技术，在有界上下文下成本降低多达 50%，Terminal-Bench 上保持甚至提升性能，并在 KernelBench 上达测试时缩放效率新水平与 SOTA 结果。其每轮节省让测试时缩放的性能-成本权衡更高效：用不到两次全上下文运行的成本，在 Terminal-Bench 上增加逾 10 个百分点。并行测试时缩放下，CliffCompaction 让 Kimi K2.6 追平 Opus 4.7，并以更低成本超过 Opus 4.6 与 GPT-5.3 Codex。关键在于只截断或丢弃内容以保持忠实，绝不改写；且从不压缩"压缩结果"——每次只作用于原始内容，丢弃此前压缩输出，防上下文漂移累积。这支撑起跨越百万 token 会话的持续学习：KernelBench 上 200 步后 CUDA 内核加速 2.23×，400 步后 3.58×，超过专门搜索算法与训练智能体。已开源与脚手架无关的 API 代理实现，可用于 Claude Code、Codex 等 harness。

原文摘要

Agents often work on complex problems that require millions of tokens of context, which necessitates compacting across sessions due to limited context windows. We develop CliffCompaction, an autocompaction technique that reduces cost by up to 50% under a bounded context while maintaining or improving performance on Terminal-Bench and achieving new levels of efficiency for test-time scaling and state-of-the-art results on KernelBench. The per-rollout savings of CliffCompaction make the performance--cost trade-off of test-time scaling more efficient, adding over 10 percentage points on Terminal-Bench for less than the cost of two full-context runs. Under parallel test-time scaling, CliffCompaction lets Kimi K2.6 match Opus 4.7, and exceed Opus 4.6 and GPT-5.3 Codex at lower cost. The key to CliffCompaction's effectiveness is that it keeps compacted information faithful by only truncating or dropping content, never rephrasing or rewriting it. We never compact a compaction---each pass operates only on original content, and prior compacted output is discarded, preventing context drift from accumulating. These properties sustain continual learning over sessions exceeding a million tokens: on KernelBench, CliffCompaction reaches CUDA kernel speedups of \(2.23\times\) after 200 steps and \(3.58\times\) after 400 steps, surpassing specialized search algorithms and trained agents despite being a general-purpose compaction technique. We open-source a scaffold-agnostic API-proxy implementation of CliffCompaction usable with Claude Code, Codex and other harnesses.

---

*自动采集于 2026-09-24*

#论文 #arXiv #Agent #小凯