acceptodds
Under review as a conference paper at ICLR 2027

Agogic: Auditing Music Tokenization under MIDI Time-Unit Sensitivity

Abstract

CLaMP-2 Fr\'echet Music Distance (FMD) is used to rank text-to-MIDI generators, but its MIDI input encodes time in arbitrary file units. Multiplying both ticks per quarter and every delta tick by two or four preserves all note times, tempo messages and non-time events in 500 real scores, yet gives FMD 35.4/78.1 between copies of the same music. We retrain seven serializers on identical 84,576 MIDI–caption slices, 80-token prefixes and Qwen3.5-0.8B settings, evaluating one fixed 500-piece reference. The raw Beat-TSD-minus-PMT FMD gap of 136.4 contracts to 14.4 under a message-preserving common 220-TPQ export with at most 1.14 ms timing change. An exact common-unit export also changes the margins. We provide a note-only diagnostic that rounds absolute times to 1 ms and writes fixed units: all 500 real and 693 valid generated exact-rescale pairs become byte-identical. Canonical FMD for PMT/Beat-TSD/PerTok/REMI/Structured is 151.6/189.4/131.1/143.0/126.3; six-comparison sampling intervals on fixed seed-0 checkpoints place Beat-TSD above PMT, PerTok and Structured below it, and leave REMI unresolved. A within-source caption shuffle yields only a 0.015 instrument-F1 difference. The note-only remedy restores invariance to the tested exact tick rescaling; neither FMD protocol is validated against listening, and prompt-level control remains unresolved.

Then back it, or bet against it.

Related papers

Open the market on this paper to see 7 more related papers.