Minseon Kim

Senior Researcher · Microsoft Research Montréal

Ph.D. in AI, KAIST

Self-improving coding agents. Safety in realistic settings.

My goal is to build continuous self-improvement cycles for coding agents: generating tasks, acting in an environment, learning from feedback, and using that experience to improve the next attempt.

In parallel, I aim to build environments that evaluate AI safety in realistic scenarios, not just isolated prompts. I want progress in agent capability to be matched by rigorous evidence of safe, reliable behavior in practical workflows.

Open to research collaboration on agent self-improvement and realistic safety evaluation. Get in touch →

Current research

Self-improving coding agents

How can an agent turn its successes and failures into the next round of useful training? I study task synthesis, learning from trajectories, and self-retrospection as components of a sustained improvement cycle.

Related work: FrogNano · Negative PI · Self-retrospection

Safety evaluation in realistic environments

How do we test safety when an agent is acting in a realistic workflow? I focus on environments that capture context, multi-step interactions, and practical constraints, exposing failure modes that static benchmarks can miss.

Related work: SecureVibe · ContextLens · Safety in fine-tuning

Latest research

Sep. 2026
FrogNano — Training a 4B coding agent via online task synthesis.
Sep. 2026
New preprints: Self-retrospection, KV-streams, SecureVibe, and ProgramDistill.
Aug. 2026
Negative PI — Learning from negative trajectories in coding agents.

Selected publications

Full record on Google Scholar →

FrogNano: Training a 4B Coding Agent via Online Task Synthesis

ArXiv 2026

M. Kim*, Z. Shi*, E. Penaloza*, C. Cui*, R. Castanyer*, M. Hashemazadeh, I. White, J. Light, J. Kim, M. Pereira, D. Moldavskaya, C. Singh, F. Vera, B. Peng, X. Yuan*, M.-A. Côté*, A. Sordoni*

PDF Twitter

How to Learn from Negative Trajectories in Coding Agents

Blog Post 2026

M. Kim, L. Caccia, D. Moldavskaya, M. Pereira, Z. Shi, F. Vera, C. Singh, M.-A. Côté, X. Yuan, A. Sordoni

Blog Post

Shockingly Simple Self-retrospection Improves Agentic Models Without RL

arXiv 2026

J. Light, C. Z. Cui, J. Kim, R. C. Castanyer, E. Penaloza, Z. Shi, A. Sordoni, M.-A. Côté, X. Yuan, M. Kim

Minseon Kim

Current research

Self-improving coding agents

Safety evaluation in realistic environments

Latest research

Selected publications

FrogNano: Training a 4B Coding Agent via Online Task Synthesis

How to Learn from Negative Trajectories in Coding Agents

Shockingly Simple Self-retrospection Improves Agentic Models Without RL

KV-streams for Efficient Compaction in Agentic Reinforcement Learning

SecureVibe: Making Vibe Coding More Secure

ProgramDistill: From Interactive Web Apps to Verifiable Reference-Guided SWE Tasks

Learning to Extract Context for Context-aware LLM Inference

Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety Routing

Align to Misalign: Automatic LLM Jailbreak with Meta-Optimized LLM Judges

Gistify! Codebase-Level Understanding via Runtime Execution

BugPilot: Complex Bug Generation for Efficient Learning of SWE Skills

BlurGuard: A Simple Approach for Robustifying Image Protection Against AI-Powered Editing

Learning to Solve Complex Problems via Dataset Decomposition

Rethinking Safety in LLM Fine-tuning: An Optimization Perspective

Exploring Sparse Adapters for Scalable Merging of Parameter Efficient Experts

MedRiskEval: Medical Risk Evaluation Benchmark of Language Models, On the Importance of User Perspectives in Healthcare Settings

Instilling Parallel Reasoning into Language Models

Enhancing Variational Autoencoders with Smooth Robust Latent Encoding

debug-gym: A Text-Based Environment for Interactive Debugging

Automatic Jailbreaking of the Text-to-Image Generative AI Systems

Optimizing Query Generation for Enhanced Document Retrieval in RAG

Protein Representation Learning by Capturing Protein Sequence‑Structure‑Function Relationship

Effective Targeted Attacks for Adversarial Self‑Supervised Learning

Generalizable Lightweight Proxy for Robust NAS against Diverse Perturbations

Language Detoxification with Attribute‑Discriminative Latent Space

Context‑dependent Instruction Tuning for Dialogue Response Generation

Meta‑Prediction Model for Distillation‑aware NAS on Unseen Datasets

Rethinking the Entropy of Instance in Adversarial Training

Lightweight Neural Architecture Search with Parameter Remapping and Knowledge Distillation

Learning Transferable Adversarial Robust Representations via Multi‑view Consistency

Consistency Regularization for Adversarial Robustness

MRI‑based classification of neuropsychiatric systemic lupus erythematosus patients with self‑supervised contrastive learning

Adversarial Self‑Supervised Contrastive Learning

Progressive Face Super‑Resolution via Attention to Facial Landmark

T1 Image Synthesis with Deep Convolutional Generative Adversarial Networks

Research experience

Senior Researcher · Microsoft Research

Postdoctoral Researcher · Microsoft Research

Research Intern · ERA–KASL AI Safety Research, University of Oxford

Research Collaboration · Theory Center, Microsoft Research Asia

Selected talks

FrogNano: Training a 4B Coding Agent via Online Task Synthesis

Learning to Extract Context for Context-aware LLM Inference

Designing Safety Systems for LLM-based Services

Learning to Extract Context for Context-aware LLM Inference

Unsupervised Context Understanding for Safer LLMs

Designing safety systems for LLM-based services

How to obtain safety effectively and efficiently

Automatic Jailbreaking of the Text-to-Image Generative AI Systems

Research notes & writing

Academic service

Education

Ph.D., Graduate School of AI · KAIST

M.S., Electrical Engineering · KAIST

B.S., Bio & Brain Engineering and Computer Science · KAIST

Contact