Takyoung Kim

Generates full-duplex dialogue data whose turn-taking behavior adapts to each scenario using a small set of human preference annotations.

Shows how multilingual prompts and multi-turn interactions can compound weaknesses in LLM safeguards against biased responses.

Recovers six physical parameters of a simulated plate reverb from one impulse response using CMA-ES followed by a targeted ternary search.

Evaluates planning agents across intermediate behaviors as well as final outcomes to reveal what drives user satisfaction throughout an interaction.

Assigns topics to coherent text segments instead of whole documents, producing cleaner topics for documents that discuss multiple themes.

Helps agents recover from ambiguous or unsupported user requests by inserting a diagnosis and recovery plan into their reasoning at test time.

Tracks how a simulated user's goal changes across turns so LLM user simulators can generate more consistently goal-aligned responses.

Combines fine-grained checks for each response with whole-dialogue pairwise comparisons to catch errors that conventional dialogue metrics miss.

Links every math-reasoning step to its supporting premises, making long chains easier for LLMs to verify and improving error detection.

Plans tailored query outlines for RAG so generated answers cover the subtopics a user wants while avoiding those they exclude.

Finds that dialogue systems often fail or pretend to help when real users make vague requests outside the systems' assumed capabilities.

Improves dialogue state tracking by selectively masking and pre-training on important entities found in the target dialogue data.

Introduces a Korean social-bias dataset spanning 72 demographic groups and shows that filtering with it reduces bias in LLM-generated text.

Provides a large Korean dataset of sensitive questions and acceptable responses to help language models discuss delicate topics more safely.

Shows that dialogue state trackers struggle when users change their minds, and that training with explicit turnback examples restores performance.

Identifies a flaw in standard dialogue-state metrics and proposes relative slot accuracy for fairer evaluation across conversation turns.

Builds an efficient pipeline that summarizes long Korean documents and reads the result aloud in a selected speaker's voice.