English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

CliffCompaction: Cost-Efficient Compaction for Million-Token Agent Sessions

Forum topic · 小凯 · 2026-09-24

Summary

CliffCompaction is an autocompaction technique for AI agents that reduces context cost by up to 50% under a bounded context window while maintaining or improving performance. Its key principle is faithfulness: content is only truncated or dropped, never rephrased, and compacted output is never re-compacted—each pass operates only on original content, preventing context drift. On Terminal-Bench, per-rollout savings make test-time scaling more efficient, adding over 10 percentage points 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. Over sessions exceeding a million tokens, it sustains continual learning on KernelBench, reaching CUDA kernel speedups of 2.23x after 200 steps and 3.58x after 400 steps, surpassing specialized search algorithms and trained agents. A scaffold-agnostic open-source API-proxy implementation works with Claude Code, Codex, and other harnesses.

Overview

  • Field: Agents
  • Authors: Trang Nguyen, Eulrang Cho, Bingqing Chen, Tim Dettmers
  • Published: 2026-09-22
  • arXiv: 2609.26779
  • Abstract (translated)

    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.

    Key Takeaways

  • Up to 50% cost reduction under bounded context without performance loss
  • Faithful compaction: truncate/drop only, never rewrite
  • No compaction-of-compaction, avoiding context drift
  • Continual learning over million-token sessions (KernelBench speedups up to 3.58x)
  • Open-source API-proxy implementation, harness-agnostic

Tags

#ai-agents#context-compaction#test-time-scaling#llm-efficiency#cuda-kernels#arxiv-paper#kernelbench#terminal-bench

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178635144