I’m Seanie Lee.
I study how to make
AI safer.

My research focuses on AI safety, responsible AI, and evaluation. I received my PhD in Artificial Intelligence from KAIST in 2026.

Portrait of Seanie Lee

Work Experience

Apple Seattle Oct 2025 – Sep 2026

Research Intern · Machine Learning Research

Developed SimSAT, a synthetic training pipeline that helped tool-calling agents generalize to unseen APIs and user states on AppWorld and OfficeBench.

Paper: Simulate to Generalize

Host: Raviteja Vemulapalli

KRAFTON Seoul Jul 2025 – Oct 2025

Research Intern

Developed ThinkSafe, which uses self-generated safety data to improve the safety of reasoning models while keeping reasoning performance competitive.

Paper: ThinkSafe

Mila Montréal Jan 2024 – Jun 2024

Research Intern

Studied whether broader red-team coverage improves LLM safety tuning and found that it helped models withstand attacks outside the training set.

Paper: Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning

Advisor: Yoshua Bengio

Collaborator: Minsu Kim

Collaborator: Moksh Jain

Collaborator: Esmeralda S. Whitammer

Visit Mila ↗

Apple Cambridge May 2023 – Sep 2023

Research Intern

Built a lightweight conversation retriever for few-shot dialogue state tracking by distilling a model trained on conversation summaries.

Paper: Effective and Efficient Conversation Retrieval for Dialogue State Tracking with Implicit Text Summaries

Host: Anders Johannsen

Collaborator: Jianpeng Cheng

National University of Singapore Singapore Jul 2022 – Sep 2022

Research Intern · Remote

Used self-distillation to regularize continued transformer pretraining on domain-specific data, improving downstream classification in the evaluated vision and language tasks.

Paper: Self-Distillation for Further Pre-training of Transformers

Advisor: Kenji Kawaguchi

Visit National University of Singapore ↗

Education

Publications

  1. Agent Training

    EACL · 2027

    Simulate to Generalize: Scaling Stateful Supervision for API-calling Agents using LLM World Models

    SimSAT uses language models to simulate stateful APIs and generate training trajectories, helping agents generalize to unseen environments on AppWorld and OfficeBench.

  2. Safety Alignment

    NeurIPS · 2026

    THINKSAFE: Self-Generated Safety Alignment for Reasoning Models

    ThinkSafe trains reasoning models on their own filtered safety examples, improving safety while keeping reasoning performance competitive.

  3. Contextual Privacy

    NeurIPS · 2026

    It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs

    SelfCI improves context-appropriate disclosure by training models to retain information needed for the task and withhold information inappropriate to share.

  4. Agent Safety

    EMNLP · 2026

    T-MAP: Red-Teaming LLM Agents with Trajectory-aware Evolutionary Search

    T-MAP evaluates safety at the level of tool execution and finds failures that text-only evaluations can miss.

  5. Reasoning Verification

    TMLR · 2026

    Rethinking Reward Models for Multi-Domain Test-Time Scaling

    A comparison across 14 domains finds that generative outcome verifiers deliver the most consistent gains among four reward-model variants.

  6. Multimodal Safety

    CVPR Findings · 2025

    HoliSafe: Holistic Safety Benchmarking and Modeling with Safety Meta Token for Vision-Language Model

    HoliSafe benchmarks safety across combinations of images and text and adds a visual guard that helps vision-language models identify harmful images.

  7. Federated Learning

    NeurIPS · 2025

    FedSVD: Adaptive Orthogonalization for Private Federated Learning with LoRA

    FedSVD reduces noise amplification in differentially private federated LoRA by refactoring aggregated updates into an adaptive orthogonal basis.

  8. Agent Distillation

    NeurIPS Spotlight · 2025

    Distilling LLM Agent into Small Models with Retrieval and Code Tools

    Agent Distillation transfers tool use as well as reasoning from large teachers, making smaller agents competitive with larger models on factual and mathematical tasks.

  9. Retrieval & Calibration

    NeurIPS · 2025

    Reliable Decision-Making via Calibration-Oriented Retrieval-Augmented Generation

    CalibRAG retrieves documents to improve the calibration of decisions informed by language models and improves accuracy across the evaluated datasets.

  10. LLM Post-Training

    NeurIPS · 2025

    Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training

    Trajectory Balance with Asynchrony learns from asynchronously collected experience to speed up language-model post-training while retaining strong task performance.

  11. Safety Guardrails

    ACL Findings · 2025

    SafeRoute: Adaptive Model Selection for Efficient and Accurate Safety Guardrails in Large Language Models

    SafeRoute reduces moderation cost by sending difficult inputs to a larger safety model and handling easier inputs with a smaller one.

  12. Speech Accessibility

    Interspeech · 2025

    Personalized Fine-Tuning with Controllable Synthetic Speech from LLM-Generated Transcripts for Dysarthric Speech Recognition

    Personalized adaptation and synthetic dysarthric speech improve speech recognition for speakers with dysarthria.

  13. Safety Guardrails

    ICLR · 2025

    HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models

    HarmAug uses synthetic training data to help compact safety guard models detect a broader range of harmful inputs.

  14. LLM Safety Evaluation

    ICLR · 2025

    Learning Diverse Attacks on Large Language Models for Robust Red-teaming and Safety Tuning

    The study finds that broader red-team coverage can improve the robustness of safety tuning against attacks outside the training set.

+ MORE − LESS Publications from 2020 to 2024
  1. Efficient Inference

    EMNLP · 2024

    Optimized Speculative Sampling for GPU Hardware Accelerators

    Parallelizing speculative sampling on GPUs reduces runtime without changing output accuracy, and an additional approximation trades a small accuracy loss for more speed.

  2. Molecular Design

    ICML · 2024

    Drug Discovery with Dynamic Goal-aware Fragment

    GEAM learns and updates a vocabulary of molecular fragments relevant to target properties, improving results on the paper’s molecule-design benchmarks.

  3. Dialogue Systems

    NAACL · 2024

    Effective and Efficient Conversation Retrieval for Dialogue State Tracking with Implicit Text Summaries

    A lightweight retriever distilled from conversation summaries improves few-shot dialogue state tracking without generating summaries at inference time.

  4. Dataset Distillation

    ICLR · 2024

    Self-Supervised Dataset Distillation for Transfer Learning

    A small synthetic dataset transfers representations from a self-supervised model trained on full unlabeled data, reducing the cost of adapting to new tasks.

  5. Neural Architecture Search

    ICLR · 2024

    DiffusionNAG: Task-guided Neural Architecture Generation with Diffusion Models

    DiffusionNAG generates neural architectures as graphs and uses task-specific predictors to steer sampling toward promising designs.

  6. Retrieval & Reasoning

    NeurIPS · 2023

    Knowledge-Augmented Reasoning Distillation for Small Language Models in Knowledge-Intensive Tasks

    KARD distills reasoning supported by retrieved evidence into small language models, improving performance on knowledge-intensive tasks.

  7. Set Learning

    ICML · 2023

    Scalable Set Encoding with Universal Mini-Batch Consistency and Unbiased Full Set Gradient Approximation

    Universal mini-batch consistency lets models process a large set in chunks while estimating the full-set training gradient without bias and with constant memory overhead.

  8. Model Watermarking

    ICML · 2023

    Margin-based Neural Network Watermarking

    Margin-based watermarking strengthens ownership verification for neural networks against attempts to replicate their functionality.

  9. Representation Learning

    ICLR · 2023

    Self-Supervised Set Representation Learning for Unsupervised Meta-Learning

    Set-SimCLR learns representations of sets from unlabeled images and improves few-shot classification over instance-level self-supervision and unsupervised meta-learning baselines.

  10. Continued Pretraining

    ICLR · 2023

    Self-Distillation for Further Pre-training of Transformers

    Self-distillation regularizes continued transformer pretraining on domain data and improves downstream performance in the studied vision and language tasks.

  11. Meta-Learning

    NeurIPS · 2022

    Set-based Meta-Interpolation for Few-Task Meta-Learning

    Meta-Interpolation uses neural set functions to generate additional training tasks and improve adaptation when few source tasks are available.

  12. Bayesian Inference

    NeurIPS · 2022

    On Divergence Measures for Bayesian Pseudocoresets

    The paper connects dataset distillation to Bayesian pseudocoresets through reverse KL and Wasserstein objectives and unifies the two views.

  13. Set Learning

    ICML · 2022

    Set Based Stochastic Subsampling

    A two-stage subsampler combines coarse global context with sequential refinement to retain the input elements most useful for a downstream task.

  14. Multilingual Learning

    ICLR · 2022

    Sequential Reptile: Inter-Task Gradient Alignment for Multilingual Learning

    Sequential Reptile aligns gradients across language tasks during meta-training to improve multilingual transfer.

  15. Question Answering

    ACL · 2021

    Learning to Perturb Word Embeddings for Out-of-distribution QA

    Learning to perturb question and context embeddings improves question-answering performance on unseen domains.

  16. Text Generation

    ICLR · 2021

    Contrastive Learning with Adversarial Perturbations for Conditional Text Generation

    Contrastive training on positive and negative perturbations improves conditional text generation by teaching models to distinguish preserving edits from misleading ones.

  17. Meta-Learning

    ICLR · 2021

    Meta-GMVAE: Mixture of Gaussian VAE for Unsupervised Meta-Learning

    Meta-GMVAE learns from unlabeled examples with a Gaussian-mixture latent space and uses that structure to adapt to new few-shot classification tasks.

  18. Question Answering

    ACL · 2020

    Generating Diverse and Consistent QA pairs from Contexts with Information-Maximizing Hierarchical Conditional VAEs

    A hierarchical generative model produces diverse question–answer pairs from unlabeled text while encouraging the questions and answers to remain consistent.

  19. Speech Processing

    Interspeech · 2020

    g2pM: A Neural Grapheme-to-Phoneme Conversion Package for Mandarin Chinese Based on a New Open Benchmark Dataset

    g2pM introduces an open benchmark of more than 99,000 sentences and a neural model for disambiguating Mandarin characters with multiple pronunciations.

Honors and Awards

Invited Talks