PromptTSS: Interactive Time Series Segmentation

Updated 7 July 2026

PromptTSS is a prompting-based framework for interactive multi-granularity time series segmentation that assigns discrete states at both coarse and fine levels.
It uses label and boundary prompts to refine segmentation during inference without retraining, leveraging bidirectional attention between time series and prompt representations.
Empirical benchmarks show that PromptTSS achieves significant accuracy gains and robust transfer learning across diverse datasets.

PromptTSS most commonly denotes a prompting-based framework for interactive multi-granularity time series segmentation, introduced as a unified model for multivariate time series whose latent states exist at multiple levels of granularity, from coarse operational phases to fine event-level behaviors (Chang et al., 12 Jun 2025). The framework addresses two problems emphasized in the literature: existing segmentation methods generally do not handle multiple granularities within one model, and they have limited adaptability to new or evolving patterns. PromptTSS answers both by using label and boundary prompts to steer a single segmentation model at inference time, so that segmentation can be refined interactively without retraining. In arXiv usage, however, the string “PromptTSS” is not entirely unique; it also appears in other contexts, including prompt optimization and as a typographical variant of PromptTTS in speech synthesis, which makes disambiguation necessary.

1. Conceptual scope and problem formulation

PromptTSS is formulated for multivariate time series segmentation. Given a length- $T$ , $d$ -channel time series $X \in \mathbb{R}^{T \times d}$ , the objective is to assign a discrete state to each time index. At a single granularity, the target sequence is $s = (s_1,\dots,s_T)$ with $s_t \in \{1,\dots,K\}$ , and boundaries are the change points where $s_t \neq s_{t-1}$ . In the multi-granularity setting, the same sequence is associated with several state sequences $g^{(\ell)}$ , ordered from coarse to fine, with distinct label spaces and boundary sets at each level (Chang et al., 12 Jun 2025).

The central design choice is to make segmentation prompt-conditioned rather than fixed. PromptTSS introduces two prompt types. Label prompts encode sparse positive and negative information about which state should or should not occur at a timestamp. Boundary prompts encode whether a change point should or should not occur. These prompts may be supplied during training to simulate interaction, or during inference to correct, refine, or constrain the model’s output.

The training target remains per-timestep state prediction. For iteration $r$ , with predicted class probabilities $\hat{y}^{(r)}_t$ and ground-truth label $s_t$ , the per-iteration objective is the cross-entropy segmentation loss

$d$ 0

and the full objective is

$d$ 1

The paper explicitly states that there is no boundary detection loss and no hierarchical consistency loss. This is important: multi-granularity behavior is learned through unified training and prompt conditioning rather than through a dedicated cross-level constraint.

2. Prompt representation and architectural design

The architecture contains three modules: a time series encoder $d$ 2, a prompt encoder $d$ 3, and a state decoder $d$ 4 (Chang et al., 12 Jun 2025). The time series encoder is a Transformer encoder with patching. Rather than operating directly on all $d$ 5 time steps, a patching operator aggregates adjacent samples into $d$ 6 tokens, reducing the effective sequence length before Transformer processing. The resulting latent representation is $d$ 7.

Prompt encoding is defined explicitly. For a target granularity with $d$ 8 states, the label prompt at time $d$ 9 is represented as $X \in \mathbb{R}^{T \times d}$ 0. The first $X \in \mathbb{R}^{T \times d}$ 1 entries form a one-hot positive label vector, and the second $X \in \mathbb{R}^{T \times d}$ 2 entries form a multi-hot negative label vector. Its embedding is

$X \in \mathbb{R}^{T \times d}$ 3

with $X \in \mathbb{R}^{T \times d}$ 4. The boundary prompt is binary, $X \in \mathbb{R}^{T \times d}$ 5, with embedding

$X \in \mathbb{R}^{T \times d}$ 6

where $X \in \mathbb{R}^{T \times d}$ 7 is a trainable lookup. The combined prompt representation is

$X \in \mathbb{R}^{T \times d}$ 8

so missing prompts contribute zero vectors rather than special tokens.

The state decoder is a two-way Transformer decoder. One layer performs prompt self-attention, cross-attention from prompts to time-series tokens, an MLP transformation on prompt states, and cross-attention from time-series tokens back to prompts. In schematic form,

$X \in \mathbb{R}^{T \times d}$ 9

followed by an MLP, and then

$s = (s_1,\dots,s_T)$ 0

After $s = (s_1,\dots,s_T)$ 1 layers, per-timestep logits are produced from the refined prompt states:

$s = (s_1,\dots,s_T)$ 2

This organization gives prompts a first-class role in inference. Rather than merely concatenating prompt tokens to the input, the decoder uses bidirectional interaction between prompt and signal representations. A plausible implication is that PromptTSS treats prompts as structured constraints over latent segmentation rather than as auxiliary metadata.

3. Training regime and interactive inference

Training is iterative and explicitly designed to simulate interactive prompt accumulation (Chang et al., 12 Jun 2025). For each input window, prompts are initialized as empty. At iteration $s = (s_1,\dots,s_T)$ 3, the model samples additional label and boundary prompts, accumulates them into the existing prompt set, re-encodes prompts, and recomputes segmentation. The loss is evaluated at every iteration, and all iterations contribute to the final training objective. The default setting uses $s = (s_1,\dots,s_T)$ 4 iterations, adding between $s = (s_1,\dots,s_T)$ 5 and $s = (s_1,\dots,s_T)$ 6 prompts per iteration.

Inference is structurally simpler. Given an input window and a set of user-provided prompts, the model encodes the patched time series, encodes prompts, applies the two-way decoder, and outputs per-timestep state probabilities. If the user adds or edits prompts, the model reruns forward inference with the updated prompt set; no parameter updates are required. This interactive loop is central to the method’s claimed adaptability to unseen or evolving patterns.

The training and preprocessing configuration is specified in detail. Chronological splits are used, with 70% train, 15% validation, and 15% test. Sliding windows have length $s = (s_1,\dots,s_T)$ 7 for most datasets and $s = (s_1,\dots,s_T)$ 8 for USC-HAD, with stride $s = (s_1,\dots,s_T)$ 9 for $s_t \in \{1,\dots,K\}$ 0 and $s_t \in \{1,\dots,K\}$ 1 for $s_t \in \{1,\dots,K\}$ 2. Patching uses patch length $s_t \in \{1,\dots,K\}$ 3 and patch stride $s_t \in \{1,\dots,K\}$ 4. The encoder has 3 Transformer layers; the decoder has 6 two-way layers; the embedding dimension is $s_t \in \{1,\dots,K\}$ 5 with dropout $s_t \in \{1,\dots,K\}$ 6. Optimization uses AdamW with learning rate $s_t \in \{1,\dots,K\}$ 7 and weight decay $s_t \in \{1,\dots,K\}$ 8, together with early stopping.

Ablation results described in the paper indicate that iterative prompting matters. Accuracy improves as the number of training iterations increases up to $s_t \in \{1,\dots,K\}$ 9, then declines at $s_t \neq s_{t-1}$ 0, which the authors attribute to over-reliance on prompts and higher computational cost. Prompt density also has a strong effect: even at 1% prompted timesteps, accuracy on USC-HAD reaches 89.93%, while 5% prompts typically exceed 95% accuracy.

4. Empirical performance on segmentation benchmarks

PromptTSS is evaluated on industrial, wearable, and activity datasets, including Pump V35/V36/V38, IndustryMG, USC-HAD, PAMAP2, ActRecTut, and MoCap (Chang et al., 12 Jun 2025). The reported metrics are Accuracy (ACC), Macro F1 (MF1), and Adjusted Rand Index (ARI). The headline result is that PromptTSS improves average accuracy by 24.49% in multi-granularity segmentation and 17.88% in single-granularity segmentation.

In the multi-granularity setting, representative results include USC-HAD with original and 2 $s_t \neq s_{t-1}$ 1 coarser labels, where PromptTSS reaches ACC 98.56, MF1 98.00, and ARI 97.19, compared with a best baseline ACC of 82.82 from PrecTime. On Pump V35 with original, 2 $s_t \neq s_{t-1}$ 2, and 4 $s_t \neq s_{t-1}$ 3 coarser labels, PromptTSS reports ACC 97.68, MF1 95.53, and ARI 95.98, versus best baseline ACC 82.41 from U-Time. On PAMAP2 with original and 2 $s_t \neq s_{t-1}$ 4 coarser labels, PromptTSS reaches ACC 99.05, MF1 98.78, and ARI 97.46, compared with best baseline ACC 61.31 from DeepConvLSTM. On IndustryMG with fine and coarse labels, PromptTSS reports ACC 96.35, MF1 77.10, and ARI 90.93, versus best baseline ACC 86.91 from MS-TCN++.

In the single-granularity setting, PromptTSS also remains strong. On MoCap it reports ACC 91.86, MF1 91.31, and ARI 78.94, versus best baseline ACC 61.35. On USC-HAD it reports ACC 96.56, MF1 95.97, and ARI 93.45, versus best baseline ACC 75.69. On Pump V35 it reports ACC 98.29, MF1 94.28, and ARI 97.82, versus best baseline ACC 91.86. On IndustryMG fine-level segmentation it reaches ACC 97.79, MF1 89.91, and ARI 95.55, narrowly exceeding the best baseline ACC 97.65.

The reported baseline set includes PrecTime, MS-TCN++, U-Time, DeepConvLSTM, and forecasting-based models adapted to segmentation, namely iTransformer-TSS and PatchTST-TSS. The paper states that, for multi-granularity evaluation, baselines are trained as separate sub-models per granularity and then combined with a post-hoc prompt-consistency selection mechanism. PromptTSS therefore differs not only in accuracy but also in using a single unified model.

5. Transfer learning, adaptability, and computational properties

A defining empirical claim for PromptTSS is its transfer performance under unseen datasets and unseen granularity levels (Chang et al., 12 Jun 2025). In transfer from Pump V35 to V36 with 5% prompts, PromptTSS reports ACC 73.91%, MF1 52.69%, and ARI 65.78%, whereas the second-best ACC is 15.51% from DeepConvLSTM. In transfer from V35 to V38, it reports ACC 76.70%, MF1 56.82%, and ARI 66.57%, with second-best ACC 28.87% from U-Time.

The paper also evaluates transfer across granularity levels within the same dataset. Training on original plus 2 $s_t \neq s_{t-1}$ 5 coarser labels and testing on 4 $s_t \neq s_{t-1}$ 6 coarser labels yields ACC 78.06%, MF1 23.87%, and ARI 68.04%, compared with second-best ACC 9.68%. Training on original plus 4 $s_t \neq s_{t-1}$ 7 coarser labels and testing on 2 $s_t \neq s_{t-1}$ 8 coarser labels yields ACC 76.37%, MF1 33.90%, and ARI 64.81%, compared with second-best ACC 9.42%. In the hardest case, training on 2 $s_t \neq s_{t-1}$ 9 and 4 $g^{(\ell)}$ 0 coarser labels and testing on original labels yields ACC 40.27%, MF1 23.32%, and ARI 31.30%, compared with second-best ACC 7.33%. These results underpin the paper’s statement of up to 599.24% improvement in transfer learning.

The efficiency argument rests on patching and on the separation between encoder and prompt-conditioned decoder. Without patching, encoder self-attention scales as $g^{(\ell)}$ 1; with patching, it scales as $g^{(\ell)}$ 2, where $g^{(\ell)}$ 3. Decoder cost is dominated by prompt self-attention $g^{(\ell)}$ 4 and cross-attention terms $g^{(\ell)}$ 5. Because evaluated window lengths are typically 256 or 512, and because patching substantially reduces $g^{(\ell)}$ 6, the runtime remains manageable. The paper’s ablations further report that PromptTSS scales well up to $g^{(\ell)}$ 7 and that larger windows improve accuracy by supplying more context.

A plausible implication is that the framework is especially suited to domains where operator feedback is sparse but high-value, such as manufacturing process monitoring and wearable-sensor activity segmentation. That implication is consistent with the benchmark selection, but the paper itself confines its formal claims to the reported datasets and settings.

The term “PromptTSS” requires disambiguation because arXiv uses nearby or identical strings for different methods.

Usage	Domain	Note
PromptTSS	Time series segmentation	Prompting-based approach for interactive multi-granularity time series segmentation (Chang et al., 12 Jun 2025)
PromptTSS / OPTS-TS	Prompt optimization	Thompson sampling-based strategy-selection mechanism inside OPTS (Ashizawa et al., 3 Mar 2025)
PromptTSS	Speech synthesis naming	Treated as a typographical variant of PromptTTS rather than a separate method (Guo et al., 2022)

In the prompt-optimization literature, the closest exact reuse is the Thompson sampling variant of OPTS, where PromptTSS refers to a multi-armed bandit mechanism for selecting prompt design strategies during prompt optimization (Ashizawa et al., 3 Mar 2025). In the speech literature, “PromptTSS” does not identify a distinct TTS method: the relevant paper is “PromptTTS,” and the data explicitly states that “PromptTSS” should be considered a typographical variant rather than a separate system (Guo et al., 2022). Similarly, “PromptST” is a spatio-temporal multi-attribute prediction framework, not PromptTSS (Zhang et al., 2023). A common misconception is therefore to treat PromptTSS as a general prompt-based forecasting or TTS label; in current arXiv usage, the distinct 2025 segmentation framework is the clearest canonical referent.

For the segmentation framework itself, the stated limitations are precise. There is no explicit boundary detection loss, no explicit cross-level hierarchical consistency objective, and performance in multi-granularity settings is lower without prompts, even though it remains above baselines. Excessive iterative training, such as $g^{(\ell)}$ 8, can lead to prompt over-reliance. The decoder also attends over all $g^{(\ell)}$ 9 prompt positions, so very long windows can incur quadratic prompt self-attention cost. The future directions named in the paper are transition from per-timestep to mask-based segmentation and temporal consistency mechanisms that allow prompt effects to persist across windows rather than being handled independently window by window (Chang et al., 12 Jun 2025).