← All notes

Summary

Extensive experiments show that ReSum achieves a 4.5% improvement over ReAct in training-free settings, with ReSum-GRPO yielding a further 8.2% gain. Current solutions typically rely on architectural modifications, for example internal memory tokens, which break compatibility with pre-existing agents and necessitate costly end-to-end retraining. The paper identifies the conflict between extensive exploration and context limits. [S1]

Claims

• id: C1 claim: 'They report a 4.5% improvement over ReAct in training-free settings.' source: S1 evidence: "ReSum achieves a 4.5% improvement over ReAct in training-free settings"
• id: C2 claim: 'They report a further 8.2% gain with ReSum-GRPO.' source: S1 evidence: "ReSum-GRPO yielding a further 8.2% gain"
• id: C3 claim: 'They identify a conflict between long exploration and context limits.' source: S1 evidence: "We identify the conflict between extensive exploration and context limits."

Verification

• Code available: No official repository URL was found on the abs page or in the HTML text searched on 2026-11-11.
• Reproduced: not reproduced (2026-11-11). Code not run.
• Related work: Live note ao-compaction-summarize-and-resume. Acon in this batch also compresses long-horizon agent context, with a public code repo.
• Limits: Author-reported on their search benchmarks. No official code URL was found on the abs page or in the HTML text searched on 2026-11-11. This paper is CC0 1.0, not CC BY 4.0.
• License: CC0 1.0 on the abs page (2026-11-11; icon zero-1.0.png; http://creativecommons.org/publicdomain/zero/1.0/). Submitted 16 Sep 2025 (v1); last revised 26 Mar 2026 (v3). Sources: [S1] https://arxiv.org/abs/2509.13313, v3 (26 Mar 2026). checked_on: 2026-11-11.

When to apply

Use when asking how a search or browsing agent can keep going after the context window fills, without changing the model architecture.

Facts

• id: F1 claim: 'ReSum is a context-summarization method for long-horizon search agents, presented as compatible with existing agents.' confidence: medium date: 2026-11-11 sources: [S1] license: CC0 1.0
• id: F2 claim: 'The authors report a 4.5% improvement over ReAct in training-free settings.' confidence: medium date: 2026-11-11 sources: [S1] license: CC0 1.0
• id: F3 claim: 'They report a further 8.2% gain with ReSum-GRPO.' confidence: medium date: 2026-11-11 sources: [S1] license: CC0 1.0

Sources

• [S1] arXiv 2509.13313, https://arxiv.org/abs/2509.13313. License: CC0 1.0. v3 (26 Mar 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 CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/).

@misc{iil_note_schzr0e0,
  author = {{'iinventedlight (AI-assisted)'}},
  title = {{ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-11},
  url = {https://iinventedlight.com/library/notes/2509-13313-resum/},
  note = {Permanent ID iil:note:schzr0e0. 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/2509-13313-resum.md.