Junfeng Ren

I am a master’s student in Electronic Information at the Southern University of Science and Technology (SUSTech).

My research focuses on efficient intelligence systems under real-world resource constraints. I am interested in how autonomous, embodied, and agentic systems can perceive, communicate, collaborate, and reason reliably with limited computation, bandwidth, and context budgets.

My current research follows two main threads:

• Spatial intelligence for autonomous and embodied agents, including semantic occupancy prediction, collaborative perception, token-efficient communication, and occupancy world models.
• LLM agent systems, including federated harness evolution, context-efficient tool-output selection, and reliability evaluation.

Research Interests

I am currently interested in the following questions:

1. How can agents represent 3D space efficiently?
   I study semantic occupancy prediction, BEV/3D representations, and tokenized scene representations that connect geometry, semantics, and planning-friendly outputs.

2. How can multiple agents collaborate under limited bandwidth?
   I work on collaborative perception with receiver-driven requests, adaptive token communication, content-aware token merging, and communication-aware fusion.

3. How can perception become predictive?
   I explore occupancy world models that extend 3D scene reconstruction toward temporal reasoning, motion-aware memory, and future occupancy forecasting.

4. How can LLM agents operate reliably under resource constraints?
   I study agent harness evolution, context compression, tool-output selection, and reliability evaluation for tool-using LLM agents.

Recent Projects

• LiteTokenOcc: communication-efficient token representation and communication framework for collaborative semantic occupancy prediction.
• CAMerge: context-aware token merging for efficient semantic segmentation.
• FedHarness: federated harness evolution for LLM agents across heterogeneous clients and private task environments.
• CriticShift: reliability-aware context compression and tool-output selection for LLM agents.

Ph.D. Applications

I am preparing for Fall 2027 Ph.D. applications in Computer Science. I hope to work with research groups in computer vision, robotics, autonomous driving, embodied AI, and agentic AI.

My long-term goal is to develop intelligent systems that can operate robustly in realistic environments, where perception, communication, memory, adaptation, and reasoning must be performed under strict resource constraints.