Zihou Ren, Byungchul Tak and Long Wang submitted REALM: Identifying Minimal Task Assets for LLM Agents to arXiv on 29 October 2026. The authors propose identifying and maintaining the resources an agent needs for a task, including resources discovered during execution. Their stated problem is excess access to files, accounts and services, which prompt injection can exploit to redirect an agent toward unrelated resources.[1]

REALM, short for Required Execution Assets Linked to a Mission, represents required assets as dependency-graph nodes and resolves their prerequisite relationships. According to the abstract, newly discovered assets enter through source-scoped selection, separating discovery from authorization while leaving the agent free to choose its computation. A REALM Builder maintains selected grants and projects Asset Files; a Partial API Manager exposes resource-bound Partial APIs to the agent inside a Task Container.[1]

The authors report results on 60 clean tasks and 560 injected cases in AgentDyn, using qwen3.5-flash as the base model. The abstract gives task-completion rates of 40.0% and 36.4%, respectively, and reports attack success rate (ASR) falling from 16.6% to 0.2%. These are the authors’ reported evaluation results; reading the abstract does not independently establish their reproducibility or applicability to another deployment.[1]

[1]: REALM: Identifying Minimal Task Assets for LLM Agents. arXiv, 29 October 2026. https://arxiv.org/abs/2610.39155

Sources

  1. REALM: Identifying Minimal Task Assets for LLM AgentsarXiv · 29 Oct 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

9d4f50825a0637962aa41eca2baa7e67e5019345c2cafb3967cdf40ffbc92110