Sangmin Bae.

Postdoctoral Researcher · KAIST AI

I study efficient machine learning through adaptive computation, efficient architectures, efficient post-training, and agentic systems.

I received my Ph.D. from KAIST, advised by Se-Young Yun. Previously, I was a research intern at Google DeepMind and Meta FAIR.

LATEST

News

SEP 2026

Four papers at NeurIPS 2026, including BASTION, LLM-as-a-Tutor, DualDrift, and Temporal Alignment Guidance.

SEP 2026

New preprint: On-Policy Reverse Distillation for weak-to-strong generalization.

AUG 2026

TAC-CoT, our work on tool-empowered implicit reasoning, at EMNLP 2026.

Education

2021 — 2025

Ph.D. in Artificial Intelligence

KAIST

Advisor: Se-Young Yun

Dissertation

2019 — 2021

M.S. in Industrial & Systems Engineering

KAIST

Advisor: Se-Young Yun

2014 — 2019

B.S. in Industrial & Systems Engineering

KAIST

Research

Adaptive computation

Early exiting, recursive Transformers, and speculative decoding.

Efficient architectures

Hybrid models, linear attention, and scalable foundation models.

Efficient post-training

On-policy distillation and learning beyond a teacher’s limits.

Agentic systems

Model routing, orchestration, and self-evolving memory systems.

Research highlights

OPRD Figure 1: successive model transfer and multi-domain consolidation, with learning curves comparing OPRD against reinforcement learning and distillation baselines.

Enlarge figure

EFFICIENT POST-TRAINING · PREPRINT 2026

On-Policy Reverse Distillation

Using a weaker teacher’s guidance to accelerate a stronger student’s own learning.

Read paper Code

Transformer–SSM hybrid architecture showing inter-layer and intra-layer combinations.

Enlarge figure

EFFICIENT ARCHITECTURES · TMLR 2026

Hybrid Architectures

Systematic analysis and design insights for combining Transformers and state-space models.

Read paper

Mixture-of-Recursions overview: router, shared recursion block, and token-wise compute allocation.

Enlarge figure

ADAPTIVE COMPUTATION · NEURIPS 2025

Mixture-of-Recursions

Shared parameters and learned routing let each token use the depth it needs.

Read paper Code

Selected Publications

All publications (33)

* Equal contribution · † Corresponding author

2026

Preprint 2026

Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

Youngrok Park^*, Sangmin Bae^*†, Hojung Jung, Jongwoo Ko, Yunseon Choi, Young Jin Kim, Pashmina Cameron, Aaron Courville, Se-Young Yun^†

Paper Code

Enlarge figure

NeurIPS 2026

BASTION: Budget-Aware Speculative Decoding with Tree-structured Block Diffusion Drafting

Soowon Oh^*, Nam Cao^*, Yujin Kim, Hojung Jung, Huzama Ahmad, Sangmin Bae^†, Se-Young Yun^†

Paper Code

BASTION method overview: block-diffusion candidates and adaptive tree construction.

Enlarge figure

TMLR 2026

Transformer-SSM Hybrid Language Models: Systematic Analysis and Design Insights

Sangmin Bae, Bilge Acun, Chien-Yu Lin, Haroun Habeeb, Seungyeon Kim, Liang Luo, Junjie Wang, Carole-Jean Wu

Paper

Enlarge figure

2025

NeurIPS 2025

Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation

Sangmin Bae^*, Yujin Kim^*, Reza Bayat^*, Sungnyun Kim, Jiyoun Ha, Tal Schuster, Adam Fisch, Hrayr Harutyunyan, Ziwei Ji, Aaron Courville^†, Se-Young Yun^†

Paper Code

Enlarge figure

ICLR 2025

Relaxed Recursive Transformers: Effective Parameter Sharing with Layer-wise LoRA

Sangmin Bae, Adam Fisch, Hrayr Harutyunyan, Ziwei Ji, Seungyeon Kim, Tal Schuster^†

Paper

Vanilla, Recursive, and Relaxed Recursive Transformer architectures, including shared layers and per-layer LoRA modules.

Enlarge figure

2024

NeurIPS 2024

Block Transformer: Global-to-Local Language Modeling for Fast Inference

Namgyu Ho^*, Sangmin Bae^*, Taehyeon Kim, Hyunjik Jo, Yireun Kim, Tal Schuster, Adam Fisch, James Thorne^†, Se-Young Yun^†

Paper Code

Block Transformer overview with embedder, block decoder, and token decoder.

Enlarge figure

2023

EMNLP 2023

Fast and Robust Early-Exiting Framework for Autoregressive Language Models with Synchronized Parallel Decoding

Sangmin Bae^*, Jongwoo Ko^*, Hwanjun Song^†, Se-Young Yun^†

Paper Code

FREE method with synchronized parallel decoding and adaptive threshold estimation.

Enlarge figure

Experience

Download CV

SEP 2025 — PRESENT

KAIST AI

Seoul, South Korea

Postdoctoral Researcher

Working with Se-Young Yun. InnoCORE-LLM Fellow, studying distillation, adaptive computation, and efficient agentic workflows.

MAY — AUG 2024

Google DeepMind

New York, United States

Research Intern

Recursive architectures for adaptive computation. Advised by Tal Schuster and Adam Fisch.

MAY — SEP 2025

Meta FAIR

Menlo Park, United States

Research Intern

Scalable, efficient hybrid language model architectures. Advised by Carole-Jean Wu and Bilge Acun.

SEP 2018 — FEB 2019

Kakao

Pangyo, South Korea

Research Intern · Recommendation Team

Deep learning and contextual multi-armed bandits for article recommendation. Advised by Kwangseob Kim.

JUL — AUG 2018

KAIST · Optimization and Statistical Inference Lab

Daejeon, South Korea

Research Intern

Multi-armed bandit algorithms, with Se-Young Yun.

DEC 2017 — JUN 2018

KAIST · Human Factors and Ergonomics Lab

Daejeon, South Korea

Research Intern

Machine learning for cross-country skiing technique classification, with Shuping Xiong.

Research projects

JAN 2026 — PRESENT

LG AI Research

Strategic sub-agents, memory systems, and hybrid architectures for agent workflows.

FEB — JUL 2026

Samsung Research

Large language model compression and inference acceleration through speculative decoding.

JUN — DEC 2025

Google Research Grant

Efficient scaling laws for reasoning in large language models. Project Manager.

NOV 2024 — SEP 2025

National AI Research Lab

Inference-efficient algorithms and neural scaling laws. Supported by IITP; Project Manager.

Earlier collaborations

KT, 2024: Efficient LLM inference algorithms.
NIER, 2023–2024: AI for particulate-matter prediction. Project Manager.
MODULABS, 2022–2024: Deep learning for respiratory sound classification. Project Manager.
KT, 2022–2023: Neural architecture search for network failure detection. Project Manager.
NAVER AI, 2022–2023: Robust federated active learning.
ETRI, 2021–2022: Federated learning on edge devices. Project Manager.
SK Hynix, 2021: Semantic segmentation for wafer process errors. Project Manager.
SK Hynix, 2020: Data-efficient unsupervised representation learning.
ETRI, 2019: Model compression for edge analysis.
Hankook Tire, 2019–2020: Compound prediction with AI and AutoML.

Teaching & academic service

Academic reviewing

ICLRICMLNeurIPSTMLRCVPRACLEMNLPAAAI

Reviewing years

CVPR 2024, 2026 · ACL 2024, 2026 · EMNLP 2026 · AAAI 2025, 2026 · ICLR 2025, 2026 · ICML 2026 · NeurIPS 2026 · TMLR.

Teaching at KAIST

Optimization for AIAI505 · Fall 2021, 2022, 2023
Machine Learning TheoryAI603 · Spring 2021, 2023
Deep Reinforcement LearningAI611 · Spring 2022

Instruction & leadership

Instructor

MetaCode · Machine Learning, 2021–2022
ForumM · Recommendation Seminar, 2022
LG-KAIST AI · Computer Vision, 2020 & 2021
Korea Blockchain Institute · Machine Learning, 2020
Samsung DS · Deep Learning, 2020

Leadership

Server Manager, KAIST AI · March 2021–February 2023
Student Leader, OSI Lab · March 2021–March 2022

Awards & Honors

2025

InnoCORE-LLM

Postdoctoral Fellowship

2025

Google Research Grant

$30,000 · Efficient scaling laws for reasoning

2025

Google Conference Scholarship Program

$3,000

2024

Samsung Humantech Paper Awards

Silver Prize · Signal Processing

2024

Google Cloud Platform Credit Award

$30,000

2022

Best Presentation Award

Korea Computing Congress

2021

Best Paper Award

Korean Artificial Intelligence Association

2019

MicroNet Challenge

Highly Compute Efficient Distinction · CIFAR-100

2017–2019

KAIST Alumni Academic Scholarship

2017

Dean’s List

KAIST College of Engineering

Patents

Efficient Decoding of Output Sequences Using Parameter Sharing

US application 19/347,594 · October 2025

Inventors

Adam Fisch, Tal Schuster, Hrayr Harutyunyan, Ziwei Ji, Seungyeon Kim, Sangmin Bae.

Toward Enhanced Representation for Federated Re-Identification by Not-True Self-Knowledge Distillation

KR granted 10-2957234-0000 · April 2026
US application 18/472,393 · September 2023

Inventors

Se-Young Yun, Seongyoon Kim, Woojin Chung, Sangmin Bae.

Federated Learning System for Performing Individual Data Customized Federated Learning, Method for Federated Learning, and Client Apparatus for Performing Same

KR granted 10-2646338-0000 · March 2024
US application 17/975,664 · October 2022

Inventors

Jaehoon Oh, Sangmook Kim, Se-Young Yun, Sangmin Bae, Jaewoo Shin, Seongyoon Kim, Woojin Chung.

System, Method, Computer-Readable Storage Medium for Federated Learning of Local Model based on Learning Direction of Global Model

KR granted 10-2934412-0000 · February 2026
US granted 12,632,784 · May 2026

Inventors

Gihun Lee, Minchan Jeong, Se-Young Yun, Sangmin Bae, Jaeyeon Ahn, Seongyoon Kim, Woojin Chung.

All Publications

Google Scholar

33 publications

* Equal contribution · † Corresponding author

2026

Preprint 2026

Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

Youngrok Park^*, Sangmin Bae^*†, Hojung Jung, Jongwoo Ko, Yunseon Choi, Young Jin Kim, Pashmina Cameron, Aaron Courville, Se-Young Yun^†

Paper Code

NeurIPS 2026

BASTION: Budget-Aware Speculative Decoding with Tree-structured Block Diffusion Drafting

Soowon Oh^*, Nam Cao^*, Yujin Kim, Hojung Jung, Huzama Ahmad, Sangmin Bae^†, Se-Young Yun^†

Paper Code

NeurIPS 2026

LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL

Yujin Kim^*, Namgyu Ho^*, Sangmin Hwang^*, Joonkee Kim, Yongjin Yang, Sangmin Bae, Seungone Kim, Jaehun Jung, Se-Young Yun^†, Hwanjun Song^†

Paper

NeurIPS 2026

DualDrift: Combining Forward and Reverse Drifts for One-Step Generative Modeling

Hojung Jung^*, Juhyeong Kim^*, Jaehyun Kwak, Boryeong Cho, Junhyeok Yang, Youngrok Park, Sangmin Bae, Se-Young Yun^†

Paper

NeurIPS 2026

Temporal Alignment Guidance: On-manifold Sampling in Diffusion Models

Youngrok Park^*, Hojung Jung^*, Sangmin Bae, Se-Young Yun^†

Paper

EMNLP 2026

TAC-CoT: Tool-Empowered Implicit Chain-of-Thought via Interleaving Explicit Tokens

Juseon Do, Sangmin Bae, Jeesu Jung^†, Hwanjun Song^†

Citation pending

TMLR 2026

Transformer-SSM Hybrid Language Models: Systematic Analysis and Design Insights

Sangmin Bae, Bilge Acun, Chien-Yu Lin, Haroun Habeeb, Seungyeon Kim, Liang Luo, Junjie Wang, Carole-Jean Wu

Paper

Interspeech 2026

MamTra: A Hybrid Mamba-Transformer Backbone for Speech Synthesis

Tan Dat Nguyen, Sangmin Bae, Joon Son Chung, Jihoon Kim^†

Paper Project

ICLR 2026

Composer: A Search Framework for Hybrid Neural Architecture Design

Bilge Acun^*, Prasoon Sinha^*, Newsha Ardalani, Sangmin Bae, Alicia Golden, Chien-Yu Lin, Meghana Madhyastha, Fei Sun, Neeraja J. Yadwadkar, Carole-Jean Wu^†

Paper

2025

Ph.D. thesis 2025

Accelerating Large Language Model Inference via Early-Exiting Algorithms

Sangmin Bae

Paper

NeurIPS 2025

Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation

Sangmin Bae^*, Yujin Kim^*, Reza Bayat^*, Sungnyun Kim, Jiyoun Ha, Tal Schuster, Adam Fisch, Hrayr Harutyunyan, Ziwei Ji, Aaron Courville^†, Se-Young Yun^†

Paper Code

TMLR 2025

VSCoDe: Visual-Augmentation Selection for Contrastive Decoding

Sihyeon Kim^*, Boryeong Cho^*, Sangmin Bae, Sumyeong Ahn^†, Se-Young Yun^†

Paper

ICML 2025

MoHAVE: Mixture of Hierarchical Audio-Visual Experts for Robust Speech Recognition

Sungnyun Kim, Kangwook Jang, Sangmin Bae, Sungwoo Cho, Se-Young Yun^†

Paper

ICLR 2025

Relaxed Recursive Transformers: Effective Parameter Sharing with Layer-wise LoRA

Sangmin Bae, Adam Fisch, Hrayr Harutyunyan, Ziwei Ji, Seungyeon Kim, Tal Schuster^†

Paper

ICLR 2025

Multi-Task Corrupted Prediction for Learning Robust Audio-Visual Speech Representation

Sungnyun Kim, Sungwoo Cho, Sangmin Bae, Kangwook Jang, Se-Young Yun^†

Paper Code

ICLR 2025

Automated Filtering of Human Feedback Data for Aligning Text-to-Image Diffusion Models

Yongjin Yang^*, Sihyeon Kim^*, Hojung Jung, Sangmin Bae, SangMook Kim, Se-Young Yun^†, Kimin Lee^†

Paper Code

Preprint 2025

Fine-tuned In-Context Learning Transformers are Excellent Tabular Data Classifiers

Felix den Breejen, Sangmin Bae, Stephen Cha, Se-Young Yun

Paper Code

2024

NeurIPS 2024

Block Transformer: Global-to-Local Language Modeling for Fast Inference

Namgyu Ho^*, Sangmin Bae^*, Taehyeon Kim, Hyunjik Jo, Yireun Kim, Tal Schuster, Adam Fisch, James Thorne^†, Se-Young Yun^†

Paper Code

SLT 2024

Learning Video Temporal Dynamics with Asymmetric Cross-Modal Attention for Robust Audio-Visual Speech Recognition

Sungnyun Kim^*, Kangwook Jang^*, Sangmin Bae, Hoirin Kim^†, Se-Young Yun^†

Paper Code

ACL 2024Oral presentation

Hard Prompts Made Interpretable: Sparse Entropy Regularization for Prompt Tuning with RL

Yunseon Choi, Sangmin Bae, Seonghyun Ban, Minchan Jeong, Chuheng Zhang, Lei Song, Li Zhao, Jiang Bian, Kee-Eung Kim^†

Paper Code

EMBC 2024

RepAugment: Input-Agnostic Representation-Level Augmentation for Respiratory Sound Classification

June-Woo Kim, Miika Toikkanen, Sangmin Bae, Minseok Kim^†, Ho-Young Jung^†

Paper

NAACL 2024

Carpe diem: On the Evaluation of World Knowledge in Lifelong Language Models

Yujin Kim, Jaehong Yoon, Seonghyeon Ye, Sangmin Bae, Namgyu Ho, Sung Ju Hwang^†, Se-Young Yun^†

Paper Code

ICASSP 2024

Stethoscope-guided Supervised Contrastive Learning for Cross-domain Adaptation on Respiratory Sound Classification

June-Woo Kim, Sangmin Bae, Won-Yang Cho, Byungjo Lee, Ho-Young Jung^†

Paper Code

2023

NeurIPS Workshop 2023

Adversarial Fine-tuning using Generated Respiratory Sound to Address Class Imbalance

June-Woo Kim, Chihyeon Yoon, Miika Toikkanen, Sangmin Bae, Ho-Young Jung^†

Paper Code

EMNLP 2023

Fast and Robust Early-Exiting Framework for Autoregressive Language Models with Synchronized Parallel Decoding

Sangmin Bae^*, Jongwoo Ko^*, Hwanjun Song^†, Se-Young Yun^†

Paper Code

Interspeech 2023

Patch-Mix Contrastive Learning with Audio Spectrogram Transformer on Respiratory Sound Classification

Sangmin Bae^*, June-Woo Kim^*, Won-Yang Cho, Hyerim Baek, Soyoun Son, Byungjo Lee, Changwan Ha, Kyongpil Tae, Sungnyun Kim^†, Se-Young Yun^†

Paper Code

CVPR 2023

Coreset Sampling from Open-Set for Fine-Grained Self-Supervised Learning

Sungnyun Kim^*, Sangmin Bae^*, Se-Young Yun^†

Paper Code

CVPR 2023

Re-thinking Federated Active Learning based on Inter-class Diversity

Sangmook Kim^*, Sangmin Bae^*, Hwanjun Song^†, Se-Young Yun^†

Paper Code

AAAI 2023Oral presentation

Self-Contrastive Learning: Single-viewed Supervised Contrastive Framework using Sub-network

Sangmin Bae^*, Sungnyun Kim^*, Jongwoo Ko, Gihun Lee, Seungjong Noh, Se-Young Yun^†

Paper Code

2022

NeurIPS 2022

Preservation of Global Knowledge by Not-True Distillation in Federated Learning

Gihun Lee^*, Minchan Jeong^*, Yongjin Shin, Sangmin Bae, Se-Young Yun^†

Paper Code

2020

NeurIPS Workshop 2020

MixCo: Mix-up Contrastive Learning for Visual Representation

Sungnyun Kim^*, Gihun Lee^*, Sangmin Bae^*, Se-Young Yun^†

Paper Code

Preprint 2020

Accurate and Fast Federated Learning via Combinatorial Multi-Armed Bandits

Taehyeon Kim^*, Sangmin Bae^*, Jin-woo Lee, Se-Young Yun^†

Paper

ICDM Workshop 2020

SIPA: A Simple Framework for Efficient Networks

Gihun Lee^*, Sangmin Bae^*, Jaehoon Oh, Se-Young Yun^†

Paper Code

Sangmin Bae.

News

Education

Ph.D. in Artificial Intelligence

M.S. in Industrial & Systems Engineering

B.S. in Industrial & Systems Engineering

Research

Adaptive computation

Efficient architectures

Efficient post-training

Agentic systems

Research highlights

Selected Publications

2026

2025

2024

2023

Experience

KAIST AI

Google DeepMind

Meta FAIR

Kakao

KAIST · Optimization and Statistical Inference Lab

KAIST · Human Factors and Ergonomics Lab

LG AI Research

Samsung Research

Google Research Grant

National AI Research Lab

Academic reviewing

Teaching at KAIST

Awards & Honors

InnoCORE-LLM

Google Research Grant

Google Conference Scholarship Program

Samsung Humantech Paper Awards

Google Cloud Platform Credit Award

Best Presentation Award

Best Paper Award

MicroNet Challenge

KAIST Alumni Academic Scholarship

Dean’s List

Patents

Efficient Decoding of Output Sequences Using Parameter Sharing

Toward Enhanced Representation for Federated Re-Identification by Not-True Self-Knowledge Distillation

Federated Learning System for Performing Individual Data Customized Federated Learning, Method for Federated Learning, and Client Apparatus for Performing Same

System, Method, Computer-Readable Storage Medium for Federated Learning of Local Model based on Learning Direction of Global Model

All Publications

2026

TAC-CoT: Tool-Empowered Implicit Chain-of-Thought via Interleaving Explicit Tokens

2025

2024

2023

2022

2020

No papers found.