Sungkyun Chang

Doctoral researcher@c4dm,
Queen Mary University of London

[arXiv Sound @ArxivSound]

Sungkyun Chang, Keshav Bhandari, Simon Dixon, Emmanouil Benetos, "Synthesis and editing of multi-instrument audio mixtures using scalar-quantised latents with MIDI Span conditioning" https://t.co/sCkMOsmLMg

[Yinghao Ma @nicolaus625]

[8/n] 👏🏻 Thanks for co-authors
@nicolaus625 , Haiwen Xia, @Xavi_Gao , Weixiong Chen, Yuxin Ye, Yuchen Yang, @mimbres1 , Mingshuo Ding, @yizhilll , @abc43992899 , Simon Dixon, Emmanouil Benetos

[Yinghao Ma @nicolaus625]

[7/n] We hope CMI-RewardBench helps push forward:

🎼 music foundation models
🧠 multimodal alignment research
🎧 AI systems that truly understand music
etc.

[Yinghao Ma @nicolaus625]

[6/n] Reward models (RMs) are crucial for:
• RLHF / RLAIF
• test-time reranking
• large-scale model alignment

🤖 Our proposed CMI-RMs enable inference-time scaling using best-of-N reranking, improves alignment scores and musical quality, and preferred by human listeners.

[Yinghao Ma @nicolaus625]

[5/n] We find even frontier multimodal LLMs struggle to judge music well.

Example accuracy on CMI-Pref:
• Gemini-3-Pro → ~65%
• Qwen3-Omni → ~60%

Our lightweight 30M parameter reward model reaches ~78% human agreement.

[Yinghao Ma @nicolaus625]

[4/n] We introduce CMI-RewardBench, a unified framework w/ 5 datasets

It evaluates reward models on following tasks
🎼 musicality scoring
📝 text–music alignment
🎤 lyric alignment
🎧 audio reference alignment
🧠 compositional multimodal reasoning

🖥Code
https://t.co/tIVp4pHhcB

[Yinghao Ma @nicolaus625]

[3/n] To address this, we build a dataset ecosystem for music reward modeling:
1️⃣ CMI-Pref-Pseudo https://t.co/aM2skQWiBt
• 110k preference pairs
• pseudo-labeled via multimodal LLM judging
2️⃣ CMI-Pref https://t.co/zcAkV4Vlhf
• 4k human annotated comparisons by 31 experts

[Yinghao Ma @nicolaus625]

[1/n]🎵 How do we evaluate AI-generated music under complex prompts (e.g. w/ lyrics & input audio etc.)?

📊 We introduce CMI-RewardBench, a unified benchmark for evaluating music reward models under Compositional Multimodal Instruction (CMI).

🧵👇 Paper: https://t.co/spHj99dblZ

[arXiv Sound @ArxivSound]

Sungkyun Chang, Simon Dixon, Emmanouil Benetos, "RUMAA: Repeat-Aware Unified Music Audio Analysis for Score-Performance Alignment, Transcription, and Mistake Detection," https://t.co/iSPDuLx3cS

[Sungkyun Chang @mimbres1]

@EzraSandzer @EzraSandzer
Thanks for testing and providing feedback on the paper's demo. However, if you look (and listen) closely, you'll see the MIDI files appear different. Here are some places in your example where there are noticeable differences. Perhaps you could give it another go.

[arXiv Sound @ArxivSound]

``ImprovNet: Generating Controllable Musical Improvisations with Iterative Corruption Refinement,'' Keshav Bhandari, Sungkyun Chang, Tongyu Lu, Fareza R. Enus, Louis B. Bradshaw, Dorien Herremans, Simon Colton, https://t.co/CATHgvzwXM

[Catslap @CatSlapToken]

Fortune favors the slappers 😼

[arXiv Sound @ArxivSound]

``YourMT3+: Multi-instrument Music Transcription with Enhanced Transformer Architectures and Cross-dataset Stem Augmentation,'' Sungkyun Chang, Emmanouil Benetos, Holger Kirchhoff, Simon Dixon, https://t.co/LuC0ZWo1JF

[Sungkyun Chang @mimbres1]

Hello, Twitter!

[Sungkyun Chang @mimbres1]

@choih51 아 귀차나요

[Oh Jaehyeuk, 오: 재혁 @JhOhi]

http://bit.ly/coSNIR 'unitxt' 실황을 느끼기에는 이 그림이 더 유리하다.

[Sungkyun Chang @mimbres1]

@HariJung Hello Hari

[Sungkyun Chang @mimbres1]

Hello World!!