Junfeng Ren

Spatial Intelligence, Efficient Transformer Inference, and LLM Agents

prof_pic.jpg

M.S. Student, SUSTech
junfengren3253@gmail.com

I am a master's student in Electronic Information at the Southern University of Science and Technology (SUSTech). My research spans spatial intelligence and world modeling, efficient Transformer inference, and LLM agent systems.

My work includes semantic occupancy prediction and collaborative perception for autonomous and embodied systems, training-free token merging and attention-level redundancy reduction for efficient Transformers, and reliable adaptation and context-efficient tool use for LLM agents.

Research Directions

01

Spatial Intelligence & World Models

Learning structured 3D representations for perception, collaboration, and prediction in autonomous and embodied systems.

Semantic Occupancy Collaborative Perception Predictive World Models
02

Efficient Transformer Inference

Designing training-free mechanisms to reduce redundant token and attention computation in Transformers.

Token Merging Content-Spatial Consistency QKV-level Merging
03

LLM Agent Systems

Studying reliable adaptation, cross-agent knowledge transfer, and context-efficient tool use in LLM agent systems.

Agent Adaptation Cross-Agent Harness Transfer Context-efficient Tool Use

Selected Research

Spatial Intelligence & World Models

Learning to Merge Tokens for Communication-Efficient Collaborative Occupancy Prediction

Collaborative semantic occupancy prediction under limited bandwidth, using receiver-driven communication and adaptive token merging to exchange compact, task-relevant scene information.

Collaborative 4D Occupancy World Models

Modeling how shared 3D occupancy evolves over time by integrating multi-agent observations, temporal context, and future occupancy prediction.

Efficient Transformer Inference

SCMerge — Content-Spatial Consistent Token Merging

Training-free token merging that combines semantic similarity with local spatial consistency to reduce redundant visual tokens while preserving scene structure.

QKV-level Token Merging

Exploring token merging within the attention pipeline to reduce redundant QKV computation beyond conventional representation-level token reduction.

LLM Agent Systems

FedHarness-RB

Reliable cross-agent transfer of executable harness improvements, using target-side validation and selective deployment to reduce negative transfer across heterogeneous LLM agents.

CriticShift

Reliability-aware tool-output selection that preserves task-critical evidence while reducing unnecessary context for tool-using LLM agents.

Research Experience

National University of Singapore

Department of Computer Science, School of Computing

Remote Research Intern · Jun 2026 – Present

Research on reliable and efficient LLM agent systems, with a focus on cross-agent harness transfer, context-efficient tool use, and reliability evaluation.

Southern University of Science and Technology

Guangdong Provincial Key Laboratory of Advanced Wireless Communication Technologies

M.S. Research · 2024 – Present

Research on spatial intelligence and efficient Transformer inference, including collaborative semantic occupancy prediction, communication-efficient perception, world modeling, and training-free token merging.

Shandong University of Science and Technology

IoT Engineering Laboratory

Research Assistant · Jul 2023 – Jan 2024

Worked on resource-constrained embedded and IoT systems, including real-time scheduling, edge computing, and robotic system development.

I am preparing for Fall 2027 Ph.D. applications. I am broadly interested in Ph.D. opportunities related to spatial intelligence and world models, efficient Transformer inference, and LLM agent systems, as well as closely related problems in computer vision, robotics, embodied AI, and efficient foundation models.