Aaron Feng

Aaron Feng

Data Science & Statistics, History Minor
@ UC San Diego

Self-evolving AI agents, concept-level memory, calibrated uncertainty, and efficient inference.

I'm an undergraduate at UC San Diego studying Data Science and Probability & Statistics.

I am interested in self-evolving AI agents capable of test-time continual learning: systems that accumulate concepts, verify what they have learned, and improve after deployment. My focus areas include concept-level memory, calibrated uncertainty, and efficient inference.

My work spans memory-augmented LLM agents, abstract reasoning, logical reasoning, AI for scientific discovery, and uncertainty quantification.

Latest Research

View All →
Current

Research Assistant @ Q-Lab

Leading concept-level memory design for ARC-AGI abstract reasoning. Engineered System-2 retrieval (+7.5% relative gain to 59.33%), built concept dataset generation pipelines, and extended to AIME math (+9.3%).

Current

Research Assistant @ Wang Lab

Developing TrustPPI (deformation stability trust signals for PPI, 0.70–0.80 AUROC). Architected heterogeneous MoE for chemical reaction prediction on 1M+ USPTO reactions with GNN encoders.

Selected Publication

View All →

ArcMemo: Abstract Reasoning Composition with Lifelong LLM Memory

Matthew Ho, Chen Si, Zhaoxiang Feng, Fangxu Yu, Yichi Yang, Zhijian Liu, Zhiting Hu, Lianhui Qin.

Runner-up, ARC Prize 2025 Paper Awards PDF Code

Featured Project

View All →

TrustPPI: Trust Signals for Protein Interactions

Model-agnostic trust framework using deformation stability to distinguish correct from incorrect PPI predictions (0.70–0.80 AUROC), outperforming generic confidence and GP uncertainty.

Structural Biology Trust/Calibration PyTorch

LLM Foundry: Scalable Training & Inference

End-to-end distributed training stack for a 1.8B-parameter language model on 8 NVIDIA B200 GPUs with pipeline parallelism, plus TPU-optimized speculative decoding for test-time reasoning.

Distributed Training B200 GPUs PyTorch