Gyuwan Kim

Gyuwan Kim

Ph.D. Candidate in Computer Science

University of California, Santa Barbara

Hi, I am a final-year Ph.D. candidate in Computer Science at UCSB, advised by Prof. Tao Yang and co-advised by Prof. Xifeng Yan, with Prof. William Wang on my committee. I am affiliated with the UCSB Search Systems Lab and UCSB NLP group. During my Ph.D., I did summer internships at Meta Superintelligence Labs, AWS AI Labs, Apple Machine Learning Research, and Microsoft Research. Before joining UCSB, I worked as a research scientist at NAVER and received my B.S. and M.S. from Seoul National University.

My research focuses on improving the efficiency and reliability of language models by developing algorithms for adaptive computation, retrieval augmentation, generation, and data augmentation, with particular interest in efficient LLM inference (e.g., KV cache compression, speculative decoding, layer skipping), retrieval (e.g., RAG, multi-vector retrieval, reranking), and long-context reasoning. I am currently on the job market for full-time Research Scientist positions. Feel free to reach out if my background aligns with your team’s mission.

I am always happy to connect with people who share similar research interests, whether for research discussions, potential collaborations, informal coffee chats, or future opportunities.

News:

  • [Aug 2026] 📄 CoinRAG has been posted on arXiv.
  • [Aug 2026] 📄 GRAFT has been posted on arXiv.
  • [Aug 2026] 🎓 Received the 2026 KSEA-KUSCO Graduate Scholarship.
  • [Aug 2026] Attended UKC 2026 in Orlando, FL. ☀️
  • [Jul 2026] Attended ICML 2026 in Seoul, South Korea. 🇰🇷
  • [Jun 2026] Passed Ph.D. proposal. 🎯
  • [May 2026] Attended CAIS 2026 in San Jose, CA.
  • [May 2026] Served as an organizer and attended KSEA IMPACTs 2026.
  • [Apr 2026] 📄 KnapSpec has been accepted at ICML 2026.
  • [Apr 2026] 📄 PPA-Plan has been accepted at ACL 2026.
  • [Jan 2026] 📄 EM-MIA has been accepted at EACL 2026.
  • [Dec 2025] Attended NeurIPS 2025 in San Diego, CA. 🌴
  • [Sep 2025] Finished a research internship at Meta Superintelligence Labs.
  • [Sep 2025] 📄 AcuRank has been accepted at NeurIPS 2025.
  • [Sep 2025] Passed Ph.D. qualifying exam and advanced to candidacy. ✨
Load more

Interests

  • Efficient LLM Inference
  • Retrieval & RAG
  • Long-Context Reasoning

Education

  • Ph.D. in Computer Science, Sep 2021 - Present

    University of California, Santa Barbara

  • M.S. in Electrical and Computer Engineering, Mar 2015 - Aug 2017

    Seoul National University

  • B.S. in Electrical and Computer Engineering / Mathematical Sciences (double major), Mar 2010 - Aug 2014

    Seoul National University

  • High School (early graduation), Mar 2008 - Feb 2010

    Seoul Science High School

Experience

 
 
 
 
 

Research Intern

Meta Superintelligence Labs

Jun 2025 – Sep 2025 Menlo Park, CA, United States

Mentors: Yixin Nie, Lin Guan, Na Zhang, and Lin Chen

  • LLM post-training to diversify responses for user engagement
 
 
 
 
 

Applied Scientist Intern

AWS AI Labs

Jun 2024 – Sep 2024 Seattle, WA, United States

Mentors: Yang Li, Evangelia Spiliopoulou, Jie Ma, and Miguel Ballesteros

  • Data contamination detection for LLMs
 
 
 
 
 

Research Intern

Apple Machine Learning Research

Jun 2023 – Sep 2023 Cupertino, CA, United States

Mentors: Awni Hannun, Angelos Katharopoulos, Edouard Grave, and Ronan Collobert

  • Efficient long-context modeling with retrieval-augmented language models
 
 
 
 
 

Research Intern

Microsoft Research

Jun 2022 – Sep 2022 Redmond, WA, United States (Remote)

Mentors: Shuohang Wang, Reid Pryzant, Yichong Xu, Yang Liu, and Chenguang Zhu

  • Efficient long-document summarization
 
 
 
 
 

Ph.D. Candidate

University of California, Santa Barbara

Sep 2021 – Present Santa Barbara, CA, United States

  • Efficient inference and adaptive computation

    • KV cache compression and reuse
    • Speculative decoding
    • Layer skipping

  • Retrieval and long-context modeling

    • Retrieval-augmented generation
    • Multi-vector retrieval
    • Reranking
    • Context construction and compression

  • Reliability and evaluation

    • Membership inference and data contamination detection
    • Source attribution for generated text
    • Reliable reasoning and planning
 
 
 
 
 

Research Scientist

NAVER Clova & AI LAB

Sep 2017 – Aug 2021 Seongnam, Republic of Korea

  • Fundamental NLP/LLM research

  • Large-scale LM pre-training/fine-tuning @ Language Representation by Clova (LaRva)

    • Memory-augmented language models
    • Data augmentation for NLP models
    • Knowledge distillation for lightweight models
    • Efficient transformer inference
    • Korean question answering ( KorQuAD)

  • AI voice agent for restaurant reservations @ Contact Center AI (CCAI)

    • Natural language understanding (intent classification and slot filling)
    • Efficient dialog state tracking
    • Automatic speech recognition error correction
    • End-to-end spoken language understanding

  • Korean grammatical error correction

  • Language model-based query auto-completion

  • LINE sticker reply recommendation

  • Community question answering based on semantic matching

 
 
 
 
 

Data Scientist Intern

Devsisters

May 2017 – Jul 2017 Seoul, Republic of Korea
  • User action modeling for churn prediction
  • Customer service automation
 
 
 
 
 

M.S. Student

Seoul National University

Mar 2015 – Aug 2017 Seoul, Republic of Korea
  • Language model-based intrusion detection system
  • RNA/protein secondary structure prediction
 
 
 
 
 

Undergraduate Research Intern

Seoul National University

Dec 2012 – Aug 2014 Seoul, Republic of Korea
  • Biological sequence classification
  • Parallel programming for biological sequence alignment

Publications

$*$ denotes eqaul contribution

CoinRAG: Contextualized Information Nugget KV Cache Reuse for Long-Context RAG

Fine-grained KV cache reuse that assembles query-relevant nugget caches from retrieved chunks through two-stage retrieval, improving the accuracy-latency Pareto frontier for long-context RAG
Paper
CoinRAG: Contextualized Information Nugget KV Cache Reuse for Long-Context RAG

Global Optimization and Inference-Time Region Grafting for Agentic Workflows

Inference-time adaptation of agentic workflows that locally replaces underperforming regions using label-free execution-quality signals, avoiding whole-workflow re-optimization
Paper
Global Optimization and Inference-Time Region Grafting for Agentic Workflows

KnapSpec: Self-Speculative Decoding via Adaptive Layer Selection as a Knapsack Problem

Self-speculative decoding that reformulates draft model selection as a knapsack problem to maximize tokens-per-time throughput, accounting for context-dependent attention overhead - ICML 2026
Paper Code
KnapSpec: Self-Speculative Decoding via Adaptive Layer Selection as a Knapsack Problem

PPA-Plan: Proactive Pitfall Avoidance for Reliable Planning in Long-Context LLM Reasoning

Proactive pitfall avoidance in plan-and-execute frameworks for long-context LLM reasoning by preventing logical pitfalls and false assumptions before plan generation - ACL 2026
Paper
PPA-Plan: Proactive Pitfall Avoidance for Reliable Planning in Long-Context LLM Reasoning

Detecting Training Data of Large Language Models via Expectation Maximization

Membership inference attack for LLMs via an expectation-maximization algorithm that handles the inherent ambiguity of membership in pretraining data - EACL 2026
Paper Code Poster Slides Video
Detecting Training Data of Large Language Models via Expectation Maximization

AcuRank: Uncertainty-Aware Adaptive Computation for Listwise Reranking

Adaptive listwise reranking that dynamically adjusts both the amount and target of computation via iterative Bayesian TrueSkill relevance estimation - NeurIPS 2025
Paper Code Poster Video
AcuRank: Uncertainty-Aware Adaptive Computation for Listwise Reranking

ScholarBench: A Bilingual Benchmark for Abstraction, Comprehension, and Reasoning Evaluation in Academic Contexts

Bilingual benchmark for evaluating LLMs on deep expert knowledge and complex academic problem-solving across abstraction, comprehension, and reasoning tasks - Findings of EMNLP 2025
Paper Code Dataset Video
ScholarBench: A Bilingual Benchmark for Abstraction, Comprehension, and Reasoning Evaluation in Academic Contexts

MacRAG: Compress, Slice, and Scale-up for Multi-Scale Adaptive Context RAG

Hierarchical RAG framework that compresses and partitions documents at multiple scales to address imprecise retrieval and fragmented context in long-context LLMs
Paper Code
MacRAG: Compress, Slice, and Scale-up for Multi-Scale Adaptive Context RAG

Towards Standardizing Korean Grammatical Error Correction: Datasets and Annotation

Three datasets from diverse sources and an automatic error-type annotation tool to standardize Korean grammatical error correction evaluation - ACL 2023
Paper Code Video
Towards Standardizing Korean Grammatical Error Correction: Datasets and Annotation

Bridging the Training-Inference Gap for Dense Phrase Retrieval

Bridging the training-inference gap for dense phrase retrieval via unified loss and hard negatives mined through efficient subcorpus validation - Findings of EMNLP 2022
Paper Poster Slides Video
Bridging the Training-Inference Gap for Dense Phrase Retrieval

Consistency Training with Virtual Adversarial Discrete Perturbation

Consistency training for text classification via virtual adversarial discrete perturbation that maximizes prediction divergence while preserving original semantics - NAACL 2022
Paper Code
Consistency Training with Virtual Adversarial Discrete Perturbation

Dynamic-TinyBERT: Boost TinyBERT's Inference Efficiency by Dynamic Sequence Length

Dynamic sequence length reduction to further enhance the inference efficiency of TinyBERT beyond static compression - ENLSP Workshop @ NeurIPS 2021
Paper Code Model Poster
Dynamic-TinyBERT: Boost TinyBERT's Inference Efficiency by Dynamic Sequence Length

NASCUP: Nucleic Acid Sequence Classification by Universal Probability

Nucleotide sequence classification using compact context-tree models and universal probability from information theory, achieving BLAST-like accuracy at much lower runtime - IEEE Access 2021
Paper Code Dataset
NASCUP: Nucleic Acid Sequence Classification by Universal Probability

SSMix: Saliency-based Span Mixup for Text Classification

Saliency-based span mixup for text classification that performs mixing on input text rather than on hidden vectors, preserving the locality of original texts - Findings of ACL 2021
Paper Code
SSMix: Saliency-based Span Mixup for Text Classification

Length-Adaptive Transformer: Train Once with Length Drop, Use Anytime with Search

Train-once, anytime-inference framework for any transformer via length drop training and multi-objective evolutionary search - ACL 2021
Paper Code Slides Video
Length-Adaptive Transformer: Train Once with Length Drop, Use Anytime with Search

Two-stage Textual Knowledge Distillation for End-to-End Spoken Language Understanding

Two-stage textual knowledge distillation from text BERT to speech model via utterance-level representation matching in pretraining and predicted logit matching in finetuning - ICASSP 2021
Paper Code Poster
Two-stage Textual Knowledge Distillation for End-to-End Spoken Language Understanding

ST-BERT: Cross-modal Language Model Pre-training For End-to-end Spoken Language Understanding

Cross-modal pretraining of speech and text for end-to-end spoken language understanding via masked and conditioned cross-modal language modeling - ICASSP 2021
Paper Poster
ST-BERT: Cross-modal Language Model Pre-training For End-to-end Spoken Language Understanding

AdamP: Slowing Down the Slowdown for Momentum Optimizers on Scale-invariant Weights

Projection of the radial component in momentum optimizers to slow down the decay of effective step sizes for scale-invariant weights - ICLR 2021
Paper Code Slides Video
AdamP: Slowing Down the Slowdown for Momentum Optimizers on Scale-invariant Weights

Large Product Key Memory for Pretrained Language Models

Large product key memory augmentation for pretrained language models with catastrophic drift mitigation for improved accuracy and speed trade-off in finetuning - Findings of EMNLP 2020
Paper Code Video
Large Product Key Memory for Pretrained Language Models

Efficient Dialogue State Tracking by Selectively Overwriting Memory

Efficient open-vocabulary dialog state tracking via selectively overwriting a fixed-sized memory with state operation prediction and slot value generation - ACL 2020
Paper Code Video
Efficient Dialogue State Tracking by Selectively Overwriting Memory

Subword Language Model for Query Auto-Completion

Subword language model for faster query auto-completion with a retrace algorithm and a reranking method by approximate marginalization - EMNLP-IJCNLP 2019
Paper Code
Subword Language Model for Query Auto-Completion

Mimicry Resilient Program Behavior Modeling with LSTM based Branch Models

Anomaly detection robust to mimicry attacks via language modeling of branch sequences - S&P 2018 DLS Workshop
Paper
Mimicry Resilient Program Behavior Modeling with LSTM based Branch Models

Training IBM Watson using Automatically Generated Question-Answer Pairs

Empirical study on training IBM Watson with automatically generated question-answer pairs as a scalable alternative to manual annotation - HICSS 2017 (IBM Best Technology Paper Honorarium)
Paper
Training IBM Watson using Automatically Generated Question-Answer Pairs

LSTM-Based System-Call Language Modelling and Robust Ensemble Method for Designing Host-Based Intrusion Detection System

System-call language modeling for anomaly-based host intrusion detection with an ensemble method to accumulate highly normal sequences and reduce false-alarm rates
Paper
LSTM-Based System-Call Language Modelling and Robust Ensemble Method for Designing Host-Based Intrusion Detection System

Academic Activities

Organizing Committee

Program Committee/Reviewer

Volunteer

  • EACL 2026, EMNLP 2022, ICML 2022, NAACL 2022, ACL 2021, NAACL 2021, ICLR 2021

Institutional Service

Presentations

Teaching

  • Teaching Assistant, Machine Learning, UCSB, Spring 2026
  • Teaching Assistant, Parallel Computing, UCSB, Winter 2026
  • Teaching Assistant, Artificial Intelligence, UCSB, Fall 2025
  • Teaching Assistant, Generative AI, UCSB, Spring 2025
  • Teaching Assistant, Problem Solving with Computers I, UCSB, Winter 2025
  • Teaching Assistant, Problem Solving with Computers II, UCSB, Fall 2024
  • Teaching Assistant, Machine Learning, UCSB, Spring 2024
  • Teaching Assistant, Machine Learning, UCSB, Winter 2024
  • Teaching Assistant, Problem Solving with Computers I, UCSB, Fall 2023
  • Teaching Assistant, Machine Learning, UCSB, Winter 2023
  • Judge, UCSB’s 10th Annual Hackathon, 2023
  • Project Mentor, NAVER CLOVA AI RUSH, 2021
  • Project Mentor, NAVER AI RUSH, 2020
  • Teaching Assistant, Machine Learning, Seoul National University, Spring 2016
  • Tutor, Programming Methodology, Seoul National University, Spring 2014
  • Problem Setter, Korean Olympiad in Informatics (KOI), 2010 – 2014
  • Student Coach, Training Camp for International Olympiad in Informatics (IOI), 2010 – 2014

Research Mentor

  • Lilliana Wang, Friendswood High School, Jun 2026 – Jul 2026
  • Joshua Shen, Westwood High School, Jun 2026 – Jul 2026
  • Hun Tae Kim, M.S. Student at UCSB, Nov 2025 – Present
  • Alan Wang, Westlake High School, Jun 2022 – Jul 2022 (now Undergrad at Carnegie Mellon University)
  • Sandra Ravishankar, Mountain View High School, Jun 2022 – Jul 2022 (now Undergrad at Duke University)
  • Soyoung Yoon, Undergrad at KAIST, Jul 2020 – Jan 2021 (now Ph.D. Student at SNU)
  • Jungsoo Park, M.S. Student at Korea University, Jul 2020 – Jan 2021 (now Ph.D. Student at Georgia Tech)
  • Seongbin Kim, M.S. Student at Inha University, Feb 2020 – Feb 2021 (now at LG Uplus)
  • Tae-Hwan Jung, Undergrad at Kyung Hee University, Dec 2019 – Jun 2020
  • Bumju Kwak, Undergrad at Seoul National University, Apr 2019 – Aug 2019 (now at Karrot)
  • Kyungwoo Song, Ph.D. Student at KAIST, Oct 2018 – Dec 2018 (now Associate Professor at Yonsei University)

Mentor

Awards

  • KSEA-SC South-Western Regional Conference 2026
    • Student Research Poster Competition (Graduate Group): 2nd Place
  • SustaiNLP Workshop 2021
    • Best Paper Award
  • HICSS 2017
    • IBM Best Technology Paper Honorarium
  • ACM International Collegiate Programming Contest (ACM-ICPC)
    • Asia Daejeon Regional: Gold Prize, 2013
    • Asia Daejeon Regional: Special Prize, 2012
    • Asia Daejeon Regional: Special Prize, 2011
  • Korean Collegiate Mathematical Competition
    • Mathematics Major Division: Silver Prize, 2013
    • Mathematics Major Division: Bronze Prize, 2012
    • Non-Mathematics Major Division: Gold Prize, 2011
  • Korea Olympiad in Informatics (KOI)
    • Gold Prize, 2008
  • Korean Mathematical Olympiad (KMO)
    • Silver Prize, 2008

Scholarships and Fellowships

Korean Computer Scientists and Engineers Association in America (KOCSEA)

  • Travel Support for KOCSEA Tech Symposium, 2025

Korean-American Scientists and Engineers Association (KSEA)

  • KSEA-KUSCO Graduate Scholarship, 2026
  • Travel Support for UKC, 2026
  • Travel Support for SEED Workshop, 2026
  • Travel Support for IMPACTs, 2026
  • Travel Support for UKC FIRE Symposium, 2025

University of California, Santa Barbara

  • Doctoral Student Travel Grant, Academic Senate, 2026
  • Zhu Family Foundation Fellowship, Graduate Division, 2026
  • GSA Conference Travel Grant, 2025
  • GSA Conference Travel Grant, 2022
  • Academic Excellence Fellowship, 2021

Seoul National University

  • Graduate Student Scholarship, 2015 – 2016
  • Partial Scholarship, 2011 – 2014

Korean Foundation for Advanced Study (KFAS)

  • Undergraduate Scholarship, 2011 – 2013

Korea Student Aid Foundation (KOSAF)

  • National Science and Engineering Scholarship, 2012