Soohan Lim

Research

My work centers on program analysis and code intelligence — using large language models (LLMs) to reason about the algorithmic properties of code and to generate specification-aware test suites. Below are the main directions I work on, with related publications listed on the Publications page.

Research map: LLM Research branches into Code Efficiency (Time Complexity Prediction → TCProF, MEC³O; Test Case Generation → STAB, ContractEval), Code Generation (Contract Enforcement → SLICE, ContractEval), Multi-Agent System (Multi-Agent Debate → MEC³O), and Korean NLP (Dialect Hate Speech → K-SpecDetox, K-ToPiC; Dialect Translation → CA-DiaL).
Research map — categories, topics, and related papers.

Time Complexity Prediction

Predicting and formally analyzing the time complexity of programs, including semi-supervised learning in low-resource settings and multi-expert consensus approaches.

Specification-driven Test Generation

Generating specification-aware test cases with LLMs — exposing algorithmic bottlenecks and checking whether generated code enforces input preconditions (contracts), often through neuro-symbolic LLM–SMT pipelines.

Contract-Satisfying Code Generation

Generating code that both implements the required functionality and enforces the input contract embedded in the problem specification — isolating contract enforcement from functional implementation through graph-based specification structuring, execution-based candidate selection, and screened input-validation assertions.

Multi-agent Debate & Reasoning

Frameworks in which multiple LLM agents debate to improve reasoning, validation, and decision-making for program analysis tasks.

Korean NLP

Detection and detoxification of Korean dialect toxicity, and dialect translation with curriculum- and attribute-based learning.