CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents
The paper claims cheaper long-context coding runs by compacting only through deletion, not rewriting.
CliffCompaction cuts per-rollout cost by up to 50% while holding or improving benchmark results. Its rule is conservative: keep original content, truncate or drop it, and never compact an already compacted summary. The authors report gains on Terminal-Bench and KernelBench, including CUDA kernel speedups after 200 and 400 steps. They say the implementation is open-source and scaffold-agnostic for Claude Code, Codex, and other harnesses. ArXiv · AI/CL/LG's note
CliffCompaction cuts per-rollout cost by up to 50% while holding or improving benchmark results. Its rule is conservative: keep original content, truncate or drop it, and never compact an already compacted summary. The authors report gains on Terminal-Bench and KernelBench, including CUDA kernel speedups after 200 and 400 steps. They say the implementation is open-source and scaffold-agnostic for Claude Code, Codex, and other harnesses. ArXiv · AI/CL/LG's note
score 7