VAIOM: Decoder-Only Financial Sequence Modeling with Continuous Inputs and Discrete Outputs

arXiv·medium signal

Yiming Ma and Xinyu Chen (arXiv 2607.13929, cs.LG/q-fin.CP) address the mismatch between decoder-only next-token architectures, which assume discrete symbolic inputs, and financial observations, which are continuous, heterogeneous, and noisy. VAIOM keeps continuous inputs while producing discrete outputs, avoiding the lossy tokenization step most financial transformer work relies on. The architectural question — how to feed genuinely continuous data to a next-token model without quantizing it first — generalizes past finance to sensor and telemetry modeling.