CV
Junfeng Ren
Shenzhen, China
Research Interests
Spatial Intelligence & World Models
Efficient Transformer Inference
LLM Agent Systems
Education
-
2024 - Present Shenzhen, China
M.S. in Electronic Information
Southern University of Science and Technology
-
2019 - 2023 Qingdao, China
B.S. in Internet of Things Engineering
Shandong University of Science and Technology
Research Experience
-
Jun 2026 – Present Singapore (Remote)
Remote Research Intern
National University of Singapore
Department of Computer Science, School of Computing
- Research on reliable and efficient LLM agent systems, with a focus on cross-agent harness transfer, context-efficient tool use, and reliability evaluation.
- Developed research systems for reliable harness adaptation and tool-output selection across heterogeneous LLM agents.
-
2024 – Present Shenzhen, China
Graduate Researcher (M.S.)
Southern University of Science and Technology
Guangdong Provincial Key Laboratory of Advanced Wireless Communication Technology
- Research on spatial intelligence and efficient Transformer inference for autonomous and embodied systems, including collaborative semantic occupancy prediction, predictive world modeling, and training-free token merging.
- Developed communication-efficient multi-agent perception and Transformer acceleration methods with an emphasis on compact token representations and redundant-computation reduction.
-
Jul 2023 – Jan 2024 Qingdao, China
Research Assistant
Shandong University of Science and Technology
IoT Engineering Laboratory
- Research on resource-constrained embedded and IoT systems, with a focus on real-time scheduling and edge intelligence.
- Conducted scheduling experiments and co-authored a peer-reviewed conference paper on RMS-based embedded scheduling.
Publications & Manuscripts
Peer-Reviewed Publications
-
Scheduling Optimization Design of IoT Embedded System Based on Improved RMS Algorithm
IEEE AIKIIE, 2023.
Manuscripts Under Review / Submitted
-
Learning to Merge Tokens for Communication-Efficient Collaborative Occupancy Prediction
Submitted to NeurIPS 2026, 2026.
-
FedHarness: Federated Harness Evolution for LLM Agents
Submitted to AAAI 2027, 2026.
-
CriticShift: Reliability-Aware Tool-Output Selection for LLM Agents
Submitted to AAAI 2027, 2026.
-
CAMerge: Context-Aware Token Merging for Efficient Semantic Segmentation
Submitted to WACV 2027, 2026.
-
Bandwidth-Aware Adaptive Token Communication for Collaborative Occupancy Prediction
Submitted to IEEE ITSC 2026, 2026.
Manuscripts in Preparation
-
Collaborative 4D Occupancy World Models with Motion-Aware Token Memory
Manuscript in preparation, 2026.
Selected Research Projects
Spatial Intelligence & World Models
-
Learning to Merge Tokens for Communication-Efficient Collaborative Occupancy Prediction
Communication-efficient collaborative semantic occupancy prediction using receiver-driven communication and adaptive token merging to exchange compact, task-relevant scene information under limited bandwidth.
-
Collaborative 4D Occupancy World Models
Predictive 3D world modeling that integrates multi-agent observations, temporal context, and future occupancy forecasting to reason about dynamic and occluded environments.
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
Exploration of 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 through target-side validation and selective deployment, reducing 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.
Selected Engineering Projects
-
Autonomous Driving Perception and Decision System on NVIDIA Jetson
Built a real-time embedded autonomous-driving prototype integrating object detection, lane perception, finite-state decision making, and vehicle control on NVIDIA Jetson.
-
Vision-Based Soccer Task System on NAO Robot
Developed a closed-loop robotic system for ball detection, navigation, alignment, and kicking using visual feedback and NAO motion control.
-
Embedded Edge-AI Face Recognition Access Control
Developed a lightweight on-device face recognition system with model compression and quantization for resource-constrained embedded deployment.