Curriculum Vitae

Education

Research Interests

• Uncertainty Quantification in LLMs: Developing methods to quantify uncertainty and confidence in LLM outputs and using these estimates to enhance reasoning performance, efficiency, and reliability.
• Learning and Adaptation in LLM Agents: Developing LLM agents that continually learn from interaction, acquire and refine reusable skills, and optimize multi-agent coordination for reliable real-world applications.

Industry Experience

• Built an AI agent that analyzes user reports to identify why content was flagged, turning noisy, large-scale, real-world feedback into reliable, actionable signals with over 90% precision.

Research Projects

• Step-Wise Confidence Attribution: Proposed a black-box framework based on the Information Bottleneck principle that assigns confidence scores to individual reasoning steps by identifying consensus structures across correct solutions, improving self-correction success rates by up to 13.5% over answer-level feedback.
• Diverse Reasoning Paths for RLVR: Studied how reasoning diversity in SFT affects downstream RL and proposed EquivAug, a symbolic-equivalence augmentation method that generates diverse, provably correct reasoning paths, achieving top post-SFT diversity and post-GRPO accuracy on GSM8K and FOLIO.
• Input Ambiguity Detection: Developed a perturbation-based framework that detects whether an input admits multiple plausible interpretations before answer generation, using response variations under controlled perturbations to distinguish inherent ambiguity from unreliable model behavior.
• Survey and Taxonomy: Authored a comprehensive survey introducing a taxonomy of UQ for LLMs based on computational efficiency and four sources of uncertainty: input, reasoning, parameters, and prediction.

• Robustness Benchmark for Mobile Agents: Introduced AndroidReality, an AndroidWorld-based benchmark testing mobile-agent robustness to state, transition, and action perturbations, and developed a training-free recovery method that improves performance in both perturbed and clean settings.
• Self-Improving Multi-Agent LLM Systems: Developed LangMARL and MASkills, two frameworks that let multi-agent LLM systems improve through interaction. LangMARL introduces agent-level language credit assignment and natural-language policy optimization; MASkills extends this to skill-level credit assignment and the continual evolution of reusable skill libraries.

Honors and Awards

• SDM Doctoral Student Forum Travel Award
• IROS 2026 IEEE RAS Travel Support
• Best Artifact Award, ICCPS 2025
• ASU Fulton Fellow Scholarship
• UBC Graduate Research Scholarship (awarded twice)
• UBC Graduate Dean’s Entrance Scholarship

Teaching

Arizona State University

Arizona State University

Service

• Tutorials: “Uncertainty Quantification and Confidence Calibration in LLMs” at ICDM 2025 and KDD 2025
• Peer Reviewer: ACL 2025 Demo, PAKDD 2026, NeurIPS 2026 Datasets & Benchmarks, ACL 2026, EMNLP 2026