Kyudan Jung

happy sunday morning. a new LongCoT king is crowned. 👑Qwen3.5-27B-Instruct + dspy.RLM yes that's right, a 27B model more than double GPT 5.2 by using recursive language models

Meet me at San Diego, ACL 2026 Industry track!

Thank you for @KAIST_AI !

🎉 Our paper " Sommelier : A Scalable Open Multi-turn Audio Pre-processing for Full-duplex Speech Language Models" is accepted at #ACL2026 industry track! We have introduced a pipeline for generating the real-world speech data necessary to build full-duplex audio language models.

We look forward to your interest in our paper and code repository! and.. meet me at San Diego~ 🍇 paper: arxiv.org/abs/2603.25750 🍄 demo: kyudan1.github.io/sommelier.gith… ✨ code: github.com/naver-ai/somme… [1] arxiv.org/abs/2407.05361

Many people have expressed curiosity regarding the name Sommelier. We chose it to honor the legacy of Emilia[1], the well-known audio data preprocessor. Our goal was to find a word that shares a similar phonetic resonance with "Emilia," ensuring that the two names feel connected.

도망간 지도교수한테 배운 쓸모 있는 몇 안되는 것들 중 하나... 흑백프린터로 뽑았을 때도 구분이 가능해야한다. 유럽애들 중에 10%정도는 색맹이다. 대가 누구누구도 색맹이다. 걔네도 보려면 색 말고도 다르게 해서 구분 가능해야한다...

Our DAVIAN Lab @KAIST_AI KAIRI intern, Seokhoon Moon and I tackled how we can enhance speech by navigating the trade-offs of three kinds of noise (additive, reverberation, and nonlinear) and proposed SLICE. Enjoy! arxiv.org/abs/2603.05302

Self-supervised speech encoders, such as WavLM, contain diverse dimensions of information. In our paper, SNAP, we claim that we can highlight artifacts to distinguish deepfake speech by nulling the speaker information in the latent representations. @KAIST_AI

Hanyang University, SNU, @KAIST_AI collaborated at NAVER Cloud! OmniACBench: A Benchmark for Evaluating Context-Grounded Acoustic Control in Omni-Modal Models arXiv: lnkd.in/gcyWR8zE

Full duplex speech dataset preprocessing paper is out! arxiv.org/abs/2603.25750

Thrilled to announce our 'Talk to your Slides' paper is accepted at #ACL2026 Findings! This paper explores how to edit PPT slides with maximum efficiency. It’s a project that truly helped me grow; after facing initial rejections,

I’m excited about the potential to build an agent system that might be even more efficient and precise than the Claude PPT agents many of you are using today. A huge thanks to all my co-authors. Meet me at San Diego! @KAIST_AI

We studied one of our recent models and found that it draws on emotion concepts learned from human text to inhabit its role as “Claude, the AI Assistant”. These representations influence its behavior the way emotions might influence a human. Read more: anthropic.com/research/emoti…