01 / Research
I’m a slow learner, but I keep learning and moving forward.
DL / Optimizer / Muon
We study the convergence and spectral learning dynamics of Simplified Muon in diffusion denoising, combining theoretical analysis of linear models with controlled experiments on nonlinear denoising networks.
LLM / Interpretability / Chain-of-Thought Reasoning
We study why Chain-of-Thought reasoning improves long-context counting across modern LLMs. Our mechanistic analysis reveals a shift from one-shot broad retrieval to iterative targeted retrieval, accompanied by more compact internal counter representations.
LLM / Evaluation / Pairwise Preference
We study fixed-confidence Pareto frontier identification from multi-objective pairwise preferences, develop an adaptive elimination algorithm with instance-dependent minimax guarantee on sample complexity, and evaluate it on synthetic benchmarks and real-world LLM Arena data.
LLM / Application / Router
We develop statistical models for prompt-dependent LLM routing from pairwise preferences, exploring prompt and model representations and trade-offs among response quality, cost, and latency.
DLM / Interpretability / Latent Reasoning
We study how diffusion language models represent and update latent reasoning states, and how these mechanisms change across architectures and decoding algorithms.
Agent / Safety / Memory Monitoring
We investigate how LLM agents maintain safety-critical information in internal activations and external memory, and whether monitoring these states can detect mistakes, overlooked requirements, or inaccurate success reports in controlled tasks.