TEMPO: Time Series Understanding via Discrete Tokenization

Time-series signals are a primary data modality in medicine, manufacturing, and engineering, yet integrating them into large language models remains an open challenge. Prior approaches either project continuous encoder representations into the LLM via cross-attention, which can collapse during training, or serialize individual timestep amplitudes as text, discarding temporal structure. We propose TEMPO, which extends an LLM's vocabulary with discrete codes from a context-aware transformer tokenizer using finite scalar quantization. Signal and text tokens share the same embedding space and self-attention stream, requiring no architectural modification and fewer than 1% trainable parameters. On five benchmarks spanning activity recognition, sleep staging, bearing-fault diagnosis, and time-series QA, TEMPO with a 4B-parameter backbone matches or exceeds cross-attention and text-serialization baselines while producing natural-language reasoning grounded in the signal.