Deprecated: The each() function is deprecated. This message will be suppressed on further calls in /home/zhenxiangba/zhenxiangba.com/public_html/phproxy-improved-master/index.php on line 456 Haebin Seong
Haebin Seonghaebin.world@gmail.com
Education
KAIST 2021.9 ~ 2023.8
Daejeon, Korea
M.S. - Artificial Intelligence
Machine Learning and Artificial Intelligence Lab (MLAI)
Seoul, Korea
under the supervision of Professor Sung Ju Hwang Sung Ju Hwang
POSTECH 2016.3 ~ 2021.8
Pohang, Korea
B.S. - Computer Science
Achieved GPA 3.70, Major GPA 3.96 (4.3 scale) Magna Cum Laude
Publications
[arXiv] The Last Harness You'll Ever Build
Haebin Seong, Li Yin, Haoran Zhang
We present a two-level automated framework that turns manual harness engineering for AI agents into an optimization problem, using a Harness Evolution Loop to improve task-specific agent harnesses and a Meta-Evolution Loop to optimize the evolution process itself across domains.
AI Agents, Harness Engineering, Automated Optimization, Meta-Learning, Agent Evaluation
[CVPR 2026 Workshop] CostNav: A Navigation Benchmark for Real-World Economic-Cost Evaluation of Physical AI Agents
Haebin Seong*, Sungmin Kim*, Yongjun Cho*, Myunchul Joe, Geunwoo Kim, Yubeen Park, Sunhoo Kim, Yoonshik Kim, Suhwan Choi, Jaeyoon Jung, Jiyong Youn, Jinmyung Kwak, Sunghee Ahn, Jaemin Lee, Younggil Do, Seungyeop Yi, Woojin Cheong, Minhyeok Oh, Minchan Kim, Seongjae Kang, Samwoo Seong, Youngjae Yu, Yunsung Lee
CostNav shows that today's navigation benchmarks miss what actually matters for real-world deployment: economic viability, not just task success. Our testbed is the first to model the full cost-revenue lifecycle of delivery robots, revealing that optimizing for navigation success often fails to optimize for commercial viability.
Physical AI, Navigation, Autonomous Delivery Systems, Robotics Economics, Economic Cost-Driven Optimization, Benchmark
[ICLR 2026] D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI
Suhwan Choi*, Jaeyoon Jung*, Haebin Seong*, Minchan Kim, Minyeong Kim, Yongjun Cho, Yoonshik Kim, Yubeen Park, Youngjae Yu, Yunsung Lee
D2E (Desktop to Embodied AI) shows that desktop interactions, especially gaming, provide scalable pretraining for robotics. Using 1.3K+ hours of data created by pseudo-labeling internet scale youtube videos, D2E achieves 96.6% on LIBERO manipulation and 83.3% on CANVAS navigation, establishing desktop pretraining as a practical robotics paradigm.
Physical AI, Vision-Language-Action (VLA) model, Embodied AI, Inverse Dynamics Model (IDM)
[ICLR 2025] HarmAug: Effective Data Augmentation for Knowledge Distillation of Safety Guard Models
Seanie Lee*, Haebin Seong*, Dong Bok Lee, Minki Kang, Xiaoyin Chen, Dominik Wagner, Yoshua Bengio, Juho Lee, Sung Ju Hwang
We propose HarmAug, a data augmentation method that distills a large safety guard model into a smaller 435M-parameter model by generating harmful instruction-response pairs using LLMs, resulting in a model that matches or outperforms larger models in F1 score and AUPRC while significantly reducing computational costs.
Large Language Model (LLM), AI Safety, LLM Guardrails, Jailbreak, AI Red-Teaming
[ACL 2025] SafeRoute: Adaptive Model Selection for Efficient and Accurate Safety Guardrails in Large Language Models
Seanie Lee*, Dong Bok Lee*, Dominik Wagner, Minki Kang, Haebin Seong, Tobias Bocklet, Juho Lee, Sung Ju Hwang
We propose SafeRoute, a binary router that distinguishes hard examples from easy ones. Our method selectively applies the larger safety guard model to the data that the router considers hard, improving efficiency while maintaining accuracy compared to solely using the larger safety guard model.
Large Language Model (LLM), AI Safety, LLM Guardrails
[NeurIPS 2025] FedSVD: Adaptive Orthogonalization for Private Federated Learning with LoRA
Seanie Lee*, Sangwoo Park*, Dong Bok Lee*, Dominik Wagner, Haebin Seong, Tobias Bocklet, Juho Lee, Sung Ju Hwang
FedSVD reduces LoRA's DP-SGD noise by using SVD reparameterization, improving stability and performance.
Differential Privacy, Federated Learning, LoRA
[AAAI 2026 Workshop] BiasJailbreak:Analyzing Ethical Biases and Jailbreak Vulnerabilities in Large Language Models
Isack Lee*, Haebin Seong*
Large language models contain hidden ethical biases that attackers can exploit to jailbreak them, and this paper both reveals those vulnerabilities and introduces a lightweight defense that reduces such bias-driven safety failures.
Large Language Model (LLM), AI Safety, Jailbreak, Social Bias, AI for Social Good
Experience
Sylph.AI 2026.03 ~ Current
(O-1A Visa) San Francisco, CA
Founding Member
Research
Building AdaL, the world's most human-aligned highest-quality AI automation lab.
Researching automated harness engineering through The Last Harness You'll Ever Build, which formalizes two-level harness and meta-evolution loops for adapting AI agents to complex workflows without manual harness engineering.
Trained customized models with pre-training and post-training methods to learn from domain-specific knowledge bases, enabling adaptive inference-time memory knowledge integration for specialized workflows.
Developing agent memory and context-management systems that support cross-session memory, memory compaction, and graph retrieval, with benchmark-driven evaluation.
Production
Established production-focused testing practices with minimal mocks and end-to-end validation to confirm each code version behaves reliably in real production conditions before release.
Built telemetry and observability workflows for maximum transparency and reproducibility, enabling customer issues to be reconstructed from a user ID within minutes without requiring additional customer context.
Built agentic proactive issue discovery across codebase, health, and log monitoring, connecting detected issues to automated product self-healing with agents.
Led core infrastructure work across LLM router reliability, MCP and OAuth integrations, cross-platform packaging, billing reliability, anti-bot measures, and traffic scaling.
Maum.ai 2025.03 ~ 2026.03
Gyeonggi, Korea
WoRV (World Model for Robotics and Vehicle control)
Senior AI Research Scientist
Pre-training and Post-training of Foundational Visual-Language-Action (VLA) Models, for autonomous driving and robotics manipulation
Sim-to-Real Digital Twin Transfer of Foundational Models to Real-World Robots with ROS, Isaac-Sim, Omnigraph, PhysX, and Omniverse
Construction of Physical Data Factory, collecting and training bi-manual humanoid teleoperation data for industry
Key Manager of Data Collection and Processing Pipeline for Physical AI, also powering D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI. Ranked #2 in Huggingface Paper of the Day.
Lead of Project CostNav: A Navigation Benchmark for Real-World Economic-Cost Evaluation of Physical AI Agents
High-Fidelity Physics Simulation with Dynamics for effective Real-World Economic Scenarios, supporting Segway E1 delivery robot, food cargo dynamics with popcorn, detailed collision dynamics, pedestrians
Real-world referenced Cost-Revenue Model with Break-Even Point Analysis, supporting Energy Cost, Pedestrian Safety Cost, Property Damage Cost, Repair Cost
Rule based and Learning-based Navigation Evaluation, supporting Nav2 with AMCL and GPS, ViNT, NoMAD, Sketchdrive, NavDP
Presented at CES 2026
Theori 2023.11 ~ 2025.03
Seoul, Korea
AI for Offensive Security
AI Engineer
Consulting for AI Safety (Red-Teaming LLMs, Blue-Teaming with Guard Models)
AI Agent framework for finding Security Vulnerabilities in Blackbox, Whitebox Web Applications
Security consultant report generation with Retrieval-Augmented Generation(RAG) and LLM fine-tuning with custom datasets
Trained custom fine-tuned models for various security tasks (Webpage Classification, PII Detection, CWE Classification, ...)
Lead of Project Xint Autopen: Offensive Security AI Engine of Xint
AI Agents for Automated Pentesting and Web Crawling
Automated Pentesting with AI Agents: IDOR, XSS, Open Redirect, Sensitive Page Detection by crawling and fuzzing blackbox web applications
Autonomous Page and Component Analysis: Automatically discover and test web pages and components for a wide range of security issues, including unauthorized access to admin pages and privileged APIs.
CTF Player
Engaged in various CTF (Capture-The-Flag) competitions, including DEFCON, as a member of teams "The Duck" and "Maple Mallard Magistrates".
Achieved 1st place in DEFCON 2024.
Kakao 2021 Summer
Gyeonggi, Korea
Recommender Systems Dept. (Internship)
Recommender Systems Researcher & Developer
Item-to-Item recommendation for KakaoStory similar post recommendation,
using content-based(CB) and collaborative filtering(CF) with multi-armed bandit(MAB)
User-to-Item recommendation for Kakaotalk birthday gift recommendation,
using LinUCB contextual bandit with click log pLSI features as the context vector
Fitmedi (Fitcare) 2019.12 ~ 2020.6
Seoul, Korea
Mobile App Full Stack Developer
Backend Developer
MicroService Architecture with seperate Auth, CRUD, Custom Logic server
GraphQL API using Hasura(PostgreSQL CRUD) & Apollo(Custom Logic)
JWT token generating Auth server with OAuth for naver, kakao, google, facebook
AWS Elastic Beanstalk & AWS Lambda(Serverless) for auto scaling and load balancing
Frontend Developer
React Native app targeting both iOS & Android
GraphQL API using Apollo Client & graphql-codegen
Designer Cooperation with Zeplin
Google play console & Apple developer page experience up to production
Wide interest in almost every CS related topic : AI, ML, Web, Security, System, Network, ...
Currently most interested in AI Agents, Physical AI, AI Safety
Certifications
IELTS General Training (International English Language Testing System) 2025.01.19
Scored 8.0 : Very Good User
"The test taker has fully operational command of the language with only occasional unsystematic inaccuracies and inappropriate usage. They may misunderstand some things in unfamiliar situations. They handle complex and detailed argumentation well."
IELTS Academic (International English Language Testing System) 2023.07.01
Scored 7.5 : Good User
"The test taker has operational command of the language, though with occasional inaccuracies, inappropriate usage and misunderstandings in some situations. They generally handle complex language well and understand detailed reasoning."
New TEPS (Test of English Proficiency developed by Seoul National University) 2020.09.19
Scored 519, percentile rank of 96.67%
"Near-native level of English proficiency. A score at this level typically indicates the highest English proficiency for a non-native speaker. A test taker at this level is able to perform technical tasks required in a specialized field after short-term training."
PDAO is a blockchain community and open-source foundation based at POSTECH, designed as a Decentralized Autonomous Organization (DAO) to promote blockchain development, research, and education.
PDAO operates through the PDAO Chain, a multi-chain governance platform, and focuses on community building, open-source projects, and integrating cryptocurrency into university operations.