Long-horizon face expensive inference and degraded performance as contexts grow to millions of tokens. CliffCompaction is an autocompaction technique that truncates or drops content when a context threshold is reached—never rephrasing or rewriting—and always compacts only the original content to prevent drift.
On Terminal-Bench, CliffCompaction reduced cost by up to 50% while maintaining or improving success rates. For , it added over 10 percentage points at less than the cost of two full-context runs. On KernelBench, it achieved CUDA kernel speedups of 2.23× after 200 steps and 3.58× after 400 steps, surpassing specialized methods. Under parallel test-time scaling, it allowed Kimi K2.6 to match Opus 4.7 and exceed Opus 4.6 and GPT-5.3 Codex at lower cost. The authors open-source the scaffold-agnostic implementation.