Curriculum Vitae

First author2026.07 – 2026.09

Task

Embedding and detecting a watermark in already generated code for provenance tracking, without hurting functional correctness and without false detections on human-written code.

Method

Code style rules collected from style guides and transformation rules are applied with style choices fixed by a secret key and each program's structural context. Detection calibrates watermark evidence with style probabilities estimated from human-written code and aggregates evidence across structurally matching locations so repeated choices cannot inflate it.

Tech

Style-rule transformations, keyed style selection, style-preference calibration, context-aware evidence aggregation; CodeContests across three LLMs and three languages

First author2026.05 – 2026.07

Task

Generating code from a natural-language specification that satisfies both the functional requirements and the input contracts.

Method

The specification is structured as a graph to separate functional requirements from contract conditions. Candidate bodies from the functional view are ranked by executing specification examples; contract assertions are generated separately, screened by execution, and merged into the selected body.

Tech

Specification graph, multi-candidate LLM generation, execution-based verification, token log-probability selection

Co-first author2026.02 – 2026.05

Task

Generating test inputs that trigger an algorithm's worst-case execution path from the natural-language specification alone, without a reference implementation.

Method

A Constraint Saturator computes the largest admissible input size from the variable relations and range constraints, and an Adversarial Scenario Injector injects input structures from 13 bottleneck scenarios.

Tech

LLM constraint interpretation, CP-SAT constraint optimization, isolated DOMjudge · Docker execution

Co-first author2025.09 – 2026.02

Task

Building a benchmark that tests whether generated code actually checks its input contracts (type, range, and structure constraints).

Method

An LLM translates reference assertions into SMT-LIB constraints, an SMT solver finds inputs that violate exactly one targeted contract, and reference-code execution keeps only tests that trigger the intended violation, yielding 364 tasks.

Tech

Neuro-symbolic pipeline (LLM + Z3), HumanEval+/MBPP+, execution-based test filtering

First author2026.01 – 2026.05

Task

Detecting profanity in Korean dialect sentences, quantifying its toxicity, and generating detoxified sentences that keep the dialect endings.

Method

Token log-probabilities are treated as a signal; FFT spectral energy and a sampling-discrepancy score feed an SVM detector. Toxicity is graded into three levels, level-specific prompts produce the detoxified sentence, and a spectral re-verification loop regenerates weak outputs.

Tech

FFT spectral analysis, SVM, dictionary-based pre-processing, LLM prompting

Co-first author2024.12 – 2025.02

Task

Predicting the time-complexity class of source code (O(1) to O(2ⁿ)).

Method

The best-performing LLM per complexity class is assigned as its expert with a class-specific instruction; experts exchange feedback in a structured debate, and a weighted consensus of expertise and confidence makes the final decision.

Tech

Multi-agent debate, expert role prompting, weighted consensus algorithm

Co-first author2025.03 – 2025.07

Task

Detecting profanity spans in Korean dialect sentences and polishing only those spans while preserving the dialect style.

Method

A 30,457-pair standard-to-dialect dictionary applied to KOLD yields a 28,435-sentence dialect profanity dataset; span detection is trained with a span-length curriculum, and the detected span plus a masked sentence guide the LLM to soften only the necessary expressions.

Tech

Curriculum learning, LLM fine-tuning (polyglot-ko, kullm-polyglot), detoxification generation (gemma, Bllossom)

Co-author2024.02 – 2024.12

Task

Predicting the time complexity of code in a low-resource setting with few labeled examples.

Method

Code augmentation (loop conversion, back-translation) and a symbolic module enrich the features, and two models trained on original and augmented data exchange pseudo-labels through co-training. I handled the baseline experiments and the co-training implementation.

Tech

Semi-supervised learning (co-training), AST · regex symbolic analysis, code augmentation

First author2024.08 – 2024.09

Task

Training an LLM translator that converts Gyeongsang dialect sentences into standard Korean.

Method

Training data are split into three curriculum levels by the number of dialect words per sentence, and speaker gender and age-group attributes are prepended as a header so the model learns easy sentences and speaker patterns first.

Tech

QLoRA fine-tuning (Solar, KoAlpaca-Polyglot, Korean-Bllossom), AI-Hub dialect corpus