Research map / Time series / Foundation models for time series
LAST SToP for Modeling Asynchronous Time Series
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LASTS, a text-prompt framework that serializes asynchronous event sequences with natural-language event names for LLMs, plus Stochastic Soft Prompting (StoP), a prefix-truncation soft prompt tuning method, evaluated on forecasting, imputation, and anomaly detection.
Problem
Modeling asynchronous time series (irregularly timed, multi-type events described in natural language) with LLMs, beyond forecasting to tasks like anomaly detection and imputation, where TPP models struggle with many event types and lose text semantics.
Approach
Serialize each event as (inter-arrival time, natural-language event name) in a prompt for a frozen-or-tuned LLM, and train soft prompts with stochastic prefix truncation (StoP) so every prefix becomes a valid prompt, yielding coarse-to-fine token structure.
Key results
- On Breakfast forecasting, LASTS + StoP reaches M-F1 0.2633 and MAE 32.55 vs. 0.2558 M-F1 for LASTS + QLoRA and 0.0240 for LLMTime.
- StoP outperforms vanilla soft prompting by an average 12.69% Macro-F1 and QLoRA by 13.55% across tasks and datasets.
- On the standard TPP benchmarks, LASTS + StoP outperforms existing TPP models on 13 of 18 evaluations and is top-2 on 17 of 18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LASTS + StoP instead of LASTS + SP | Llama-3-8B-Instruct on textual asynchronous time series · Forecasting and imputation across Breakfast, MultiTHUMOS, and EPIC-KITCHENS | Macro-F1: Breakfast forecasting 0.2341 -> 0.2633; MultiTHUMOS forecasting 0.3707 -> 0.3947; EPIC-KITCHENS forecasting 0.0780 -> 0. | Helped |
| Use LASTS + StoP instead of LASTS + SP | Llama-3-8B-Instruct on textual asynchronous time series · Anomaly detection across Breakfast, MultiTHUMOS, and EPIC-KITCHENS | Macro-F1: 0.6520 -> 0.7198; 0.5231 -> 0.6045; 0.6159 -> 0.6603 | Helped |
| Replace event descriptions with uniquely mapped gibberish names (ablation) | LASTS Zero Shot · Breakfast forecasting | Macro-F1: 0.0432 -> 0.0140 | Hurt |
| Represent each event as (event, time) instead of (time, event) (ablation) | LASTS Zero Shot · Breakfast forecasting, imputation, and anomaly detection | Macro-F1: 0.0432 -> 0.0409; 0.0248 -> 0.0071; 0.0760 -> 0.0858 | Mixed |
| Use durations instead of inter-arrival times (ablation) | LASTS Zero Shot · Breakfast forecasting, imputation, and anomaly detection | Macro-F1: 0.0432 -> 0.0600; 0.0248 -> 0.0140; 0.0760 -> 0.0767 | Mixed |
| Randomly select prompt tokens instead of selecting a structured prefix (ablation) | StoP prompt training · Validation data; task and dataset not specified in text | Macro-F1 | Hurt |
Surprising findings
- Converting event names to scrambled gibberish reduces zero-shot forecasting M-F1 by 67.63%, suggesting LLMs use world knowledge of event names.
- Using durations instead of inter-arrival times helps forecasting but hurts imputation and anomaly detection.
Limitations
- TPP models with explicit time distributions outperform LASTS on inter-arrival time prediction for some datasets (e.g., Amazon gap).
- Baselines like TPP models failed to converge or ran out of memory on datasets with many event types.
Experimental setup
Datasets: Breakfast MultiTHUMOS EPIC-KITCHENS-100 Amazon Retweet Taxi Taobao StackOverflow
Benchmarks: Breakfast MultiTHUMOS EPIC-KITCHENS-100 Amazon Retweet Taxi Taobao StackOverflow
Metrics: Macro-F1 MAE RMSE Accuracy
Baselines: Random Chronos LLMTime LLMProcesses RMTPP NHP SAHP THP AttNHP QLoRA
Builds on: prefix tuning QLoRA LLMTime Matryoshka representations EasyTPP benchmark
Closest papers by meaning
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.88)
- TimeCAP: Learning to Contextualize, Augment, and Predict Time Series Events with Large Language Model Agents (AAAI 2025 · similarity 0.88)
- AutoTimes: Autoregressive Time Series Forecasters via Large Language Models (NeurIPS 2024 · similarity 0.87)
- Can LLMs Understand Time Series Anomalies? (ICLR 2025 · similarity 0.87)
- Position: What Can Large Language Models Tell Us about Time Series Analysis (ICML 2024 · similarity 0.87)
- Timer: Generative Pre-trained Transformers Are Large Time Series Models (ICML 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.