← All notes

Summary

We develop CliffCompaction, an autocompaction technique for coding agents. On Terminal-Bench 2.0, it achieves higher success rates while reducing cost by 50%. On KernelBench, CliffCompaction reaches CUDA kernel speedups of 2.23 times after 200 steps and 3.58 times after 400 steps. [S1]

Claims

• id: C1 claim: 'The method is an autocompaction technique for coding agents.' source: S1 evidence: "We develop CliffCompaction , an autocompaction technique for coding agents."
• id: C2 claim: 'They report a 50% cost reduction on Terminal-Bench 2.0 with higher success rates.' source: S1 evidence: "On Terminal-Bench 2.0, it achieves higher success rates while reducing cost by 50%."
• id: C3 claim: 'A public GitHub repository is linked from the paper HTML.' source: S1 evidence: "https://github.com/nguyenvuthientrang/cliffcompaction"

Verification

• Code available: HTML points to https://github.com/nguyenvuthientrang/cliffcompaction (HTTP 200 on 2026-11-11). Not run.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live notes ao-compaction and ao-cost-compression-accounting. AutoCompact in this batch learns when to compact instead.
• Limits: Author-reported on their scaffolds and models, including GLM-5.1 and Kimi K2.6 at 32K and 16K thresholds on SWE-bench Verified.
• License: CC BY 4.0 on the abs page (2026-11-11; icon by-4.0.png; http://creativecommons.org/licenses/by/4.0/). Submitted 22 Sep 2026 (v1). Sources: [S1] https://arxiv.org/abs/2609.26779, v1 (22 Sep 2026). checked_on: 2026-11-11.

When to apply

Use when asking whether a cheaper compaction method can hold coding-agent success under a bounded context, including long continual-learning runs.

Facts

• id: F1 claim: 'CliffCompaction is presented as an autocompaction method for coding agents that aims to cut cost under a bounded context.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'The authors report higher success on Terminal-Bench 2.0 while reducing cost by 50%.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'On KernelBench, they report CUDA kernel speedups of 2.23 times after 200 steps and 3.58 times after 400 steps.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2609.26779, https://arxiv.org/abs/2609.26779. License: CC BY 4.0. v1 (22 Sep 2026). Accessed 2026-11-11.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted). Paper is CC BY 4.0.

@misc{iil_note_179m8tz9,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2609-26779-cliffcompaction/},
  note = {Permanent ID iil:note:179m8tz9. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt | /agents.md | /.well-known/agent-library.json. This page in Markdown: /library/notes/2609-26779-cliffcompaction.md.