Perception for Large Language Model-Based Agents: A Closed-Loop Framework in Open Environments

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Large language models are rapidly evolving from single-turn text generators into agentic systems capable of operating across graphical user interfaces, video streams, code execution traces, and embodied physical environments. In such contexts, reliable performance hinges upon a stable and closed-loop perceptual process that can be continuously verified and corrected over time. Although previous research has made significant progress on individual components including data acquisition, result verification, and memory management, the system-level interactions among these components remain insufficiently understood. To bridge this gap, we propose the Closed-Loop Framework for Agentic Perception in Open Environments, a unified conceptual structure that organizes the existing literature around five sequential stages: Multimodal Data Acquisition, Boundary Assessment, Perception Strategy Planning, Result Verification, and Memory Preservation. For each stage, we systematically summarize its primary patterns, representative methods, and remaining technical challenges. Building upon this framework, we survey agentic perception along five key dimensions, namely multimodal collection, boundary recognition, information scheduling, trustworthy verification, and long-term feedback, thereby distinguishing it from conventional perception paradigms in terms of openness, proactivity, and continuous self-correction capability. Finally, we identify four open challenges for future research, including unified evaluation protocols, trustworthy source modeling, cross-modal consensus mechanisms, and long-term memory maintenance, and further outline promising directions for deploying general-purpose agents in real-world applications.