PhD candidate at UCLA, graduating June 2027

Building multimodal models that are capable, efficient, robust, and fast.

I’m Wenhan, a computer science PhD student at UCLA advised by Baharan Mirzasoleiman. I work across the life of a multimodal foundation model: curating the data it learns from, post-training it with SFT, RL, and distillation, defending it against poisoning and backdoor attacks, and scaling its training and test-time inference at lower compute and latency. Lately that has extended to agentic systems, where I train lightweight orchestrators that decompose long-horizon tasks, delegate to sub-agents, and extend the effective execution horizon at a fraction of frontier-model cost.

Sixteen patches, the way a vision transformer would see it.

News

    Selected work

    1. +14.3%
      average accuracy on held-out tasks, surpassing the teacher model

      V2T: Distilling visual reasoning into text space

      Cisco Research, 2025

      Models that “think with images” are expensive at inference. V2T distills interleaved visual–textual chain-of-thought into a 7B text-only student, using logit distillation, attention alignment, and a GRPO stage rewarded for correctness and visual grounding. Training runs up to 42× faster than state-of-the-art baselines.

    2. Early
      reasoning quality as a signal for data curation

      Reasoning quality emerges early: Data curation for reasoning models

      ICML 2026

      Reasoning quality emerges early enough to guide which data a reasoning model should learn from. This work turns that early signal into a data-curation strategy for reasoning-model training.

    3. 0%
      targeted poisoning attack success, down from 93.75%

      SafeCLIP: Pre-training CLIP against targeted data poisoning and backdoor attacks

      ICML 2024

      Poisoning a tiny fraction of CLIP’s web-scale data is enough to hijack it. SafeCLIP splits training data into safe and risky sets and grows the safe set over time, breaking both targeted poisoning and backdoor attacks without hurting clean performance.

    4. 2.60×
      lossless decoding speedup across 10 multimodal benchmarks, against 2.19× for EAGLE-3

      HAWK: Rethinking multimodal drafting for speculative decoding

      Samsung Research America, 2026

      A target-aware drafter for vision-language models. HAWK grounds its drafts in the target’s compressed visual states and selected hidden layers, and a draft-aware training loss keeps it aligned with the target over multi-step drafting. Average acceptance length goes up 23%.

    Experience

    1. Jun – Sep 2026

      Samsung Research America

      Research Scientist Intern, Visual Display Intelligence Lab

      • Proposed HAWK, a multimodal drafter for lossless speculative decoding in VLMs.
      • 2.60× lossless speedup (vs. 2.19× for EAGLE-3) and 23% longer acceptance across 10 benchmarks.
    2. Jan – Jun 2026

      Google Research

      Student Researcher, Algorithms & Optimization

      • Post-trained small language models to plan over long-horizon agentic trajectories.
      • Built compact planners to replace much larger models and costly search in multi-agent systems, extending to multi-turn diagnostic agents.
    3. Jun – Sep 2025

      Cisco Research, Outshift

      Research Scientist Intern, AI & ML

      • Proposed V2T, distilling interleaved visual–textual reasoning into a text-only 7B student.
      • +14.3% average accuracy on held-out tasks and +2.7% on a broad suite, with up to 42× faster training.
    4. Jun – Sep 2024

      Samsung Research America

      Research Scientist Intern, NLU & On-Device AI Lab

      • Extended the Matryoshka Query Transformer to SALMONN with a learned, input-adaptive audio-token budget.
      • One speech LLM that runs at any token count, for efficient on-device inference.

    Publications

    * Equal contribution

    Background

    Education

    2022 – 2027
    Ph.D. in Computer Science, UCLAAdvisor: Baharan Mirzasoleiman; expected June 2027
    2019 – 2022
    B.S. Mathematics of Computation, UCLAMinor in Statistics; Magna Cum Laude; GPA 3.94

    Teaching

    CS 260D
    Large-Scale Machine LearningFall 2023, Fall 2025, Fall 2026
    CS M148
    Introduction to Data ScienceWinter 2023, Winter 2025
    CS 188
    Introduction to Human-Computer InteractionSpring 2025

    Reviewing

    Conferences
    NeurIPS, ICML (2023–26); ICLR (2023, 2026–27); AISTATS, UAI (2025–26)
    Journals
    Transactions on Machine Learning Research
    Workshops
    AdvML-Frontiers at ICML 2023 and NeurIPS 2024

    Awards and leadership

    2025
    Summer Mentored Research Award, UCLA
    2023 – 2025
    President, DanceSport Club at UCLA