Paper: frozen AMT encoders beat from-scratch training for rhythm game chart generation
kastnerkyle · x · 2026-10-15
An arXiv paper (2610.14955) by Sojeong An, Chris Donahue et al. generates rhythm game charts by freezing a pre-trained audio encoder — an automatic music transcription (AMT) model — and training only a linear projection plus Transformer decoder, since most chart times are note onsets that AMT models locate well. The approach fits human charts more closely than from-scratch encoders and places note types more accurately than MERT. Joint training on three osu! modes lifts event F1 on the lowest-data mode from 0.492 to 0.506, while transfer from Clone Hero lowers it to 0.482 — cross-game transfer depends on charting style and music coverage.
More from Research
- AlphaNIST: one RL policy solves all 20 NIST assembly tasks with 98.2% success in simulation — hbouammar · 2026-10-15
- Linear Depth Memory: matrix-valued memory tops PreLN and mHC on DepthBench across scales — SonglinYang4 · 2026-10-15
- Neural field planning for hybrid manufacturing cuts estimated build time 68.3% — jiqizhixin · 2026-10-15
- Tev-Gym: 1,170 dense-reward RL environments push long-horizon agentic training — RulinShao · 2026-10-15
- Claude systematically debunks the speculative EDAR-selenium hypothesis about East Asians — teortaxesTex · 2026-10-15
- Microsoft's AutoEnvScaling lets agents build their own RL environments, lifting Qwen3 Terminal-Bench from 41.1 to 53.3 — omarsar0 · 2026-10-15