CV

Education

  • Sep 2024 – Present

    Santa Clara, CA

    Ph.D. in Computer Science and Engineering
    University of California, Santa Cruz (Silicon Valley Campus)
    Advised by Prof. Ian Lane
    • Research interests: multimodal conversational interaction, full-duplex speech language models (SLMs), LLM hallucination, and speech processing.
    • Current research: developing full-duplex SLMs for real-time interaction, and investigating architectures for turn-taking, barge-in interruption, and latency-aware speech generation.
  • Aug 2022 – May 2024

    Durham, NC

    M.Eng. in Electrical and Computer Engineering
    Duke University
    • Coursework: Computer Systems, Systems Programming, NLP, Robust Server Software, Quantum Computing.
    • Honors: Full Scholarship.
  • Aug 2018 – May 2022

    Durham & Kunshan

    B.S. in Data Science
    Duke University & Duke Kunshan University
    Degree conferred by Duke University
    • Coursework: Computer Architecture, Computer Vision, Complex Analysis, Topological Data Analysis, Signature Capstone.
    • Honors: Dean's List with Distinction (Spring 2020); Outstanding Peer Tutor.

Industry Experience

  • Jun 2026 – Sep 2026

    Mountain View, CA

    Research Intern
    Samsung Research America
    • Proposed LI-RAG (under review at ICLR 2027) to answer complex multi-hop questions accurately at low inference cost. Strong multi-hop systems call an LLM at every retrieval step, which is slow; LI-RAG instead fine-tunes an embedding model (NV-Embed-v2 with LoRA) to imitate how an LLM breaks a question into sub-questions, so it retrieves evidence hop by hop with no LLM calls.
    • Outperformed all retrievers of comparable cost on 4 multi-hop QA benchmarks, beating the graph-based HippoRAG 2 by up to 8.6 accuracy points, and matched the LLM-driven IRCoT on most benchmarks while retrieving ~9× faster, using only 4K training questions and no human-labeled evidence.
    • Built a trainable graph retriever that re-weights the knowledge graph per query, rather than searching one fixed graph for all queries, as HippoRAG 2 does; implemented differentiable Personalized PageRank to train the edge scorer end to end.
  • Jun 2025 – Aug 2025

    San Ramon, CA

    Natural Language Processing Intern
    Five9
    • Developed two multi-agent architectures: a finite state machine flow and a stack-based coordination system, with hierarchical state management, sub-agent invocation, and dynamic tool generation, enabling seamless orchestration and context consistency across multi-turn dialogues.
    • Developed a ReAct (Reasoning + Acting) framework for both single- and multi-agent systems, integrating structured planning, reasoning, and action cycles, dynamic tool selection, and multi-step problem decomposition with iterative reasoning loops.
    • Integrated RAG-based knowledge retrieval and API simulation, including context-aware search, dynamic function creation, and knowledge group management, grounding agent reasoning in external data for reliable performance.
    • Built an automated customer simulator with dynamic persona generation and adaptive system prompts, producing realistic user behaviors and responses for robust testing of agent performance.
    • Developed a modular evaluation platform supporting both real-time and post-conversation analysis, with extensible metrics (response quality, latency, tool usage correctness, etc.) and automated report generation for comprehensive system assessment.
  • May 2022 – Aug 2022

    Beijing, China

    Machine Learning Engineer Intern
    Xiaomi Technology Corporation
    AI Lab
    • Initiated a groundbreaking project, utilizing multi-channel microphone arrays to successfully achieve multi-directional speech extraction in in-car call scenarios.
    • Developed an automated in-vehicle scene data simulation framework, resolved data scarcity issue. This provided one-click data generation, incorporating varied acoustic parameters and creating a comprehensive multi-channel, multi-speaker, multi-scene dataset.
    • Determined in-vehicle microphone array schemes. Using neural network-based beamforming estimation to construct a multi-directional speaker extraction system that effectively eliminated noise and interfering voices, as validated by subjective testing.

Research Experience

  • May 2023 – Aug 2023

    Durham, NC

    Research Assistant
    Duke University
    Advised by Prof. Jeffrey Krolik
    • Advanced speech intelligibility for the hearing-impaired through proposing and implementing a customized speech enhancement model specifically for hearing-aids. This innovative solution addressed a long-standing limitation of conventional hearing-aids.
  • Jun 2019 – May 2022

    Kunshan, China

    Research Assistant, Deep Learning-Based Target Speech Separation and Speech Recognition
    Speech and Multimodal Intelligent Information Processing (SMIIP) Lab, Duke Kunshan University
    Advised by Prof. Ming Li
    • Developed an original Automatic Speech Recognition (ASR) system using Python and NumPy, integrating MFCC extraction, dynamic time warping, HMM, and GMM for precise speech template matching.
    • Collaborated with Xiaomi researchers to extract and integrate target pitch into the separation model, pioneering two joint training strategies that improved the Target Speaker Separation (TSS) model.
    • Achieved a 5.85 dB Signal-to-Distortion Ratio (SDR) improvement (21% relative increase over baseline) on the LibriSpeech dataset by designing an LSTM-based pitch extraction model integrated with MultiBlock RNNoise (MBRNN) via joint training.
    • Secured competitive summer research funding in 2021, and contributed to a published paper at Odyssey 2022.

Projects

  • Jul 2026
    DuplexTutorBench: A Benchmark for Full-Duplex Spoken AI Tutors
    Accepted at IEEE SLT 2026, first author
    • Built the first benchmark for full-duplex spoken AI tutors, in which any real-time speech model (GPT Realtime, Gemini Live, PersonaPlex, MoshiVis) tutors a simulated student on MathVista math problems.
    • Designed a two-track evaluation suite that scores speech quality with Audiobox Aesthetics, compares turn-taking timing against human conversation (Map Task corpus) using word timings from Whisper and forced alignment, and rates teaching with an LLM judge on 8 pedagogical dimensions. Benchmarking showed that no current model does both: real-time APIs teach well but take turns unnaturally, while full-duplex models time their turns much more like humans but teach poorly.
  • Dec 2025
    TEXT2SPACE: Enhancing LLM Reasoning Capabilities via Structured Output Learning
    Accepted at COLM 2026
    • Co-developed “TEXT2SPACE”, a novel spatial reasoning benchmark that utilizes ASCII visualizations to decouple representation failures from reasoning errors, identifying a “Read-Write Asymmetry” where LLMs interpret spatial layouts effectively but struggle to reconstruct them.
    • Fine-tuned Qwen3-30B models using LoRA and Hugging Face TRL on NVIDIA A40 GPUs, implementing “Learning to Draw” (text-to-ASCII) objectives to enforce the internalization of verifiable spatial “cognitive maps” during training.
    • Demonstrated reasoning transfer by validating that auxiliary training on structured spatial construction generalizes to improve downstream text-only reasoning accuracy by ~9% and enhances performance on external benchmarks like STEPGAME.
  • Jan 2025
    Multilingual Hallucination Detection in LLM Outputs (SemEval-2025 Task 3: Mu-SHROOM)
    SemEval 2025, The 19th International Workshop on Semantic Evaluation (Best Paper Award)
    • Co-developed a system that pinpoints hallucinated spans in LLM answers across 14 languages: it retrieves evidence through web search, uses an LLM to flag unsupported facts, and maps each fact back to its exact character span in the answer.
    • Ranked #1 on average among 40+ teams and top-2 in 11 of 14 languages, after tuning prompts automatically with MiPROv2 and combining multiple systems.
  • Mar 2022
    IEEE ICASSP Signal Processing Cup 2022: Synthetic Speech Attribution
    • Proposed an end-to-end multi-system fusion solution with ResNet backbone and ranked #6 among 25+ teams.
    • Performed Gaussian disturbance with MATLAB and train/dev mismatching strategy to augment data for robustness.