Reliable External Strategy Memory for Frozen Language Models
2026
Collaborative research project · Submitted to AAAI 2027 (under review)
I framed recurring cascade failures as opportunities to build reusable external knowledge, distilling verified interventions into compact, source-free strategy cards to avoid weight updates and repeated teacher calls.
I designed a failure-triggered control flow: the frozen student first answers independently, retrieves memory after a failure, and calls the teacher only if the memory-assisted retry also fails.
Multimodal Beam Prediction for Wireless Communication Systems
2025–2026
Undergraduate research project · SUSTech
I built a 64-class mmWave vehicle-to-infrastructure beam-prediction pipeline on DeepSense 6G Scenario 31, combining camera, LiDAR, radar, GNSS, and previous power measurements with modality-specific encoders and Transformer-based fusion.
I implemented train-only, position-aligned static memory: Laplace-smoothed beam distributions are indexed by 2D GNSS cells, and the retrieved prior is fused with a frozen five-modal backbone.