AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

Hierarchical Procedural Meta-Reasoning for Generalizable Multimodal Agents

Yao Fu, Shengyi Qian, Pierluca D'Oro, Fanyi Xiao, Honglak Lee, Joseph Tighe, Manchen Wang

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

The paper proposes hierarchical procedural meta-reasoning to improve generalization of multimodal agents on complex tasks such as mobile navigation; only the abstract was available.

Problem

Multimodal agents fine-tuned for tasks like mobile navigation generalize poorly across diverse applications, frequent system changes, and customized workflows, which the authors attribute to insufficient task-specific procedural knowledge.

Approach

Abstract indicates a hierarchical procedural meta-reasoning approach intended to supply agents with procedural knowledge for generalization; specific mechanism not stated in the available text.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
condition the actor on a static procedure generated oncePLM-8B actor · DigiData-Bench-Auto (all)success rate (%): 41.47 -> 46.30Improved
update the procedure online with internal knowledge (ablation)PAMAR-Static · DigiData-Bench-Auto (all)success rate (%): 46.30 -> 52.10Improved
add prompt-based meta-reasoner with selective search (ablation)PAMAR-Dynamic · DigiData-Bench-Auto (all)success rate (%): 52.10 -> 54.20Improved
add selective external search via meta-reasoner on unseen apps (ablation)PAMAR-Dynamic · DigiData-Bench-Auto (cross-app)success rate (%): 40.51 -> 47.06Improved
always search instead of selective search (ablation)PAMAR with PLM-8B · DigiData-Bench-Auto (all)success rate (%): 54.20 -> 54.90No clear effect
add static procedure on updated app versions (ablation)PLM-8B actor · DigiData-AppUpdatesuccess rate (%): 29.50 -> 30.35No clear effect
allow dynamic procedure revision on updated app versions (ablation)PAMAR-Static · DigiData-AppUpdatesuccess rate (%): 30.35 -> 55.75Improved
add trained meta-reasoner with update notes (ablation)PAMAR-Dynamic · DigiData-AppUpdatesuccess rate (%): 55.75 -> 63.93Improved
add static procedure on multi-app tasks (ablation)PLM-8B actor · DigiData-MultiAppsuccess rate (%): 0.0 -> 0.0No clear effect
add dynamic procedure and meta-reasoner on multi-app tasks (ablation)PLM-8B actor · DigiData-MultiAppsuccess rate (%): 0.0 -> 15.0Improved
delay the first predicted search trigger (ablation)PAMAR (Train) · DigiData-Bench-Auto (all)success rate (%): 53.47 -> 48.61Hurt
train the actor with augmented procedure paraphrases and abstraction instead of step-wise procedures only (ablation)PLM-8B actor with dynamic procedures · DigiData-Bench-Auto (all)success rate (%): 48.30 -> 52.10Improved

Limitations

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy