Researchers introduce CliffCompaction, an autocompaction technique designed to reduce costs for long-horizon coding agents by up to 50% under a bounded context. The method maintains or improves performance on Terminal-Bench and achieves state-of-the-art results on KernelBench by keeping compacted information faithful through truncation rather than rephrasing.

  • CliffCompaction adds over 10 percentage points on Terminal-Bench for less than the cost of two full-context runs.
  • Under parallel test-time scaling, it allows Kimi K2.6 to match Opus 4.7 and exceed Opus 4.6 and GPT-5.3 Codex at lower cost.
  • The technique prevents context drift by discarding prior compacted output and operating only on original content.
  • On KernelBench, it reaches CUDA kernel speedups of $2.23 imes$ after 200 steps and $3.58 imes$ after 400 steps.
  • A scaffold-agnostic API-proxy implementation is open-sourced for use with Claude Code, Codex, and other harnesses.

CliffCompaction makes the performance-cost trade-off of test-time scaling more efficient, sustaining continual learning over sessions exceeding a million tokens.