TEMPO

Overview

Across domains, experts turn raw temporal signals into reasoned interpretations: reading temporal patterns, identifying which features matter, and producing conclusions grounded in the signal, the operating context, and accumulated domain knowledge. A classifier cannot replicate this process: it selects from a closed set and discards both the reasoning structure and the contextual information that expert judgment depends on.

LLMs are a natural fit: their output space is language and their pretraining encodes substantial domain knowledge. The difficulty is conditioning the model on raw numerical signal rather than relying on textual metadata. Prior approaches use continuous encoder embeddings as soft prompts (cross-attention gates can collapse during training) or scalar binning per timestep (discards temporal structure). Following the simplification arc in vision–language models, TEMPO extends the LLM vocabulary with discrete codes from a context-aware tokenizer. Signals are processed by standard self-attention alongside text.

Three design choices

Three choices distinguish TEMPO from prior signal–LLM integrations and together yield a parameter-efficient framework that transfers across modalities.

1. 01
   Vocabulary · Discrete extension

   Signals as LLM tokens

   625 quantised codes plus two delimiters are appended to the LLM vocabulary; signal and text tokens interleave in a single flat sequence and are processed by standard self-attention. No cross-attention, no gating, no architectural change to the backbone.

2. 02
   Tokenizer · Context-aware FSQ

   Transformer encoder, no codebook collapse

   A 4-layer Transformer encoder applies self-attention over the full signal before quantisation, so the same local pattern maps to different codes depending on surrounding context. Finite scalar quantisation replaces VQ-VAE and eliminates collapse; the 625-code codebook reaches 100% utilization.

3. 03
   Training · Parameter-efficient

   <1% of parameters trained

   Two stages: embedding alignment for the 627 new tokens (LLM frozen), then unified multi-task fine-tuning with DoRA adapters (rank 32). Backbone weights and original text embeddings stay frozen throughout. The full pipeline costs under $100 in compute.

Headline results

Across activity recognition, sleep staging, bearing-fault diagnosis, UCR classification, and the TSAQA QA benchmark, TEMPO matches or exceeds prior fine-tuned and cross-attention baselines while producing natural-language outputs, a capability label-only classifiers cannot provide.

TS-LLM = best published time-series-aware LLM (OpenTSLM variants). LLM = best standard LLM with text serialisation, either zero-shot (GPT-4o) or fine-tuned (LLaMA3.1-8B, Qwen3-8B, Tokenized Llama-3.2). n/a indicates no published baseline of that kind for that benchmark.

Specialised classifiers remain superior on narrow tasks (Random Forest on HAR at 96%, WDCNN on CWRU at 99.2%), as expected for purpose-built models trained on identical features. TEMPO's advantage is not accuracy on a single benchmark but framework generality: the same tokenizer and training recipe adapt across modalities and produce reasoning rather than labels.

Context-dependent reasoning

On CWRU, the same vibration signal is interpreted differently depending on textual context. Given an "aged bearing, 18 months in service" prompt the model diagnoses outer-race fault; given a "newly installed non-OEM part, 2 days old" prompt with the identical signal tokens, it diagnoses normal break-in operation. Recommended actions also escalate with ISO vibration severity zone: zone A (good) returns "no immediate action"; zone D (danger) returns "schedule replacement immediately". These behaviours are structurally unavailable to label-only classifiers or post-hoc explanation pipelines.

Cross-domain code sharing

The 625-code codebook learns a taxonomy of temporal primitives, not domain-specific shapes. Across six held-out domains (ECG, Financial, Forecasting, Synthetic, UCR, UEA), pairwise cosine similarity of per-domain code-activation vectors reveals interpretable structure: ECG and UCR share codes via quasi-periodic transients (0.78); Financial and Forecasting cluster via shared trend primitives (0.73). All similarities exceed a permutation null by > 44σ. One tokenizer transfers without modification across accelerometer, EEG, ECG, and vibration data.

A competitive TS-LLM for under $100

Toward generative time-series reasoning

Because signal tokens share the same vocabulary and attention stream as text, the LLM can generate time-series codes as part of its output. This opens modalities beyond classification and QA: autoregressive forecasting via future code generation, counterfactual reasoning ("what if the fault worsened?") via conditional code synthesis, and imputation via masked code infilling, all within a single architecture and with no decoder head swap.

Getting started

Citation

@misc{forgis2026tempo,
  title  = {TEMPO: Time Series Understanding via Discrete Tokenization},
  author = {Forgis},
  year   = {2026},
  note   = {ICML 2026 Workshop}
}