← All notes

Summary

Acon (also written ACON) optimizes context compression for long-horizon agents through compression-guideline optimization and compressor distillation. The authors report a 26-54% reduction in peak token usage on AppWorld, OfficeBench, and Multi-objective QA, with better task success than existing compression baselines, and up to 46% performance improvement for smaller models by reducing context distraction. [S1]

Claims

• id: C1 claim: 'They report a 26-54% reduction in peak token usage.' source: S1 evidence: "Acon reduces peak token usage by 26–54%"
• id: C2 claim: 'They report up to 46% performance improvement for smaller models.' source: S1 evidence: "achieving up to 46% performance improvement by mitigating context distraction"
• id: C3 claim: 'They state that code is available at a public Microsoft repository.' source: S1 evidence: "Our code is available at https://github.com/microsoft/acon"

Verification

• Code available: Paper: https://github.com/microsoft/acon (HTTP 200 on 2026-11-11). Not run.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live note ao-compaction-summarize-and-resume. ReSum in this batch is a related summarize-and-continue method for search agents.
• Limits: Author-reported on the named benchmarks. The 46% figure is for smaller models in their setup, not a general guarantee.
• 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 1 Oct 2025 (v1); last revised 1 Jun 2026 (v3). Comments: ICML 2026. Sources: [S1] https://arxiv.org/abs/2510.00615, v3 (1 Jun 2026). checked_on: 2026-11-11.

When to apply

Use when asking how to compress long agent histories while keeping task success, including when a smaller model is the agent.

Facts

• id: F1 claim: 'Acon reduces peak token usage by 26-54% on AppWorld, OfficeBench, and Multi-objective QA, while improving task success over existing compression baselines, according to the authors.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F2 claim: 'They report up to 46% performance improvement for smaller models by mitigating context distraction.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0
• id: F3 claim: 'The authors state that their code is available at https://github.com/microsoft/acon.' confidence: medium date: 2026-11-11 sources: [S1] license: CC BY 4.0

Sources

• [S1] arXiv 2510.00615, https://arxiv.org/abs/2510.00615. License: CC BY 4.0. v3 (1 Jun 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_ceh7sn6b,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{ACON: Optimizing Context Compression for Long-horizon LLM Agents}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2510-00615-acon/},
  note = {Permanent ID iil:note:ceh7sn6b. 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/2510-00615-acon.md.