Research map / Time series / Foundation models for time series
BRIDGE: Bootstrapping Text to Control Time-Series Generation via Multi-Agent Iterative Optimization and Diffusion Modeling
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 BRIDGE, a text-controlled diffusion model for time-series generation that combines LLM-refined text templates (via a multi-agent iterative optimization system) with learned semantic prototypes, evaluated on 12 in-domain and 2 unseen-domain datasets.
Problem
Controlled time-series generation across domains requires instance-level control and generalization to unseen domains, but paired text-time series data is scarce and text and continuous time series are mismatched in granularity.
Approach
A multi-agent LLM framework (manager, two teams with planner/scientist/engineer/observer roles, inter-group discussion) iteratively refines dataset-agnostic text templates, which are populated per series; a diffusion model is conditioned on both these texts and prototype weights from a Prototype Assignment Module that captures domain features.
Key results
- On 12 in-domain datasets, BRIDGE achieves best MDD on most datasets, e.g., Wind MDD 0.316 vs TimeVAE 0.943 and TimeGAN 1.115.
- Refined text reduces zero-shot LSTPrompt MAE on AirPassenger from 15.12 (initial text) to 12.39, and on Sunspots from 49.71 to 42.37.
- On unseen domains (5/10-shot), BRIDGE attains the best K-L on Web (0.905 / 0.876) and competitive MDD on Stock.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add refined text instead of initial text | LLMTime zero-shot forecasting · AirPassenger | MAE: 49.36 -> 40.94 | Helped |
| add refined text instead of initial text | LLMTime zero-shot forecasting · Sunspots | MAE: 59.88 -> 48.64 | Helped |
| add textual input (ablation) | BRIDGE text-controlled generation · Electricity | J-FTSD: 1.821 -> 0.538 | Helped |
| add semantic prototypes (ablation) | BRIDGE text-controlled generation · Electricity | J-FTSD: 1.164 -> 0.538 | Helped |
| add textual input (ablation) | BRIDGE text-controlled generation · Stock | J-FTSD: 8.178 -> 7.483 | Helped |
| add semantic prototypes (ablation) | BRIDGE text-controlled generation · Stock | J-FTSD: 7.687 -> 7.483 | Helped |
| use multiple collaborating teams rather than a single micro-detail team (ablation) | text refinement for zero-shot forecasting · AirPassenger | MAE: 44.27 -> 40.94 | Helped |
| use multiple collaborating teams rather than a single macro-level team (ablation) | text refinement for zero-shot forecasting · AirPassenger | MAE: 42.57 -> 40.94 | Helped |
| remove instance context from text (ablation) | LLMTime zero-shot forecasting · AirPassenger | MAE: 40.94 -> 41.96 | Mixed |
| remove background information from text (ablation) | LLMTime zero-shot forecasting · AirPassenger | MAE: 40.94 -> 44.63 | Hurt |
| remove statistical context from text (ablation) | LLMTime zero-shot forecasting · AirPassenger | MAE: 40.94 -> 44.01 | Hurt |
| remove pattern descriptions from text (ablation) | LLMTime zero-shot forecasting · AirPassenger | MAE: 40.94 -> 44.36 | Hurt |
Surprising findings
- Concise text outperforms overly detailed text, which can mislead the model; rule-based contextual information may add confusion.
- Without text conditioning, BRIDGE w/o Text sometimes performs better in J-FTSD on Rain and Air, and larger pre-trained LLMs change performance only slightly.
Limitations
- Text templates are derived from only two datasets and rely on manual verification to exclude dataset-specific details.
- Experiments are univariate (channel-independent), limiting multivariate applicability.
- Text-control benefits are mixed; w/o Text sometimes wins on J-FTSD and MDD for some datasets.
Experimental setup
Datasets: Electricity Solar Wind Traffic Taxi Pedestrian Air Temperature Rain NN5 Fred-MD Exchange
Benchmarks: Electricity Solar Wind Traffic Taxi Pedestrian Air Temperature Rain NN5 Fred-MD Exchange
Metrics: Marginal Distribution Discrepancy K-L divergence J-FTSD human evaluation rank MAE MSE SMAPE MASE OWA
Baselines: TimeVQVAE TimeGAN GT-GAN TimeVAE Time-LLM LLM4TS TEMPO GPT4TS
Builds on: TimeDP ReAct LSTPrompt LLMTime Time-MMD
Closest papers by meaning
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.86)
- TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster (NeurIPS 2025 · similarity 0.86)
- STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis (AAAI 2025 · similarity 0.86)
- TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts (AAAI 2025 · similarity 0.86)
- VerbalTS: Generating Time Series from Texts (ICML 2025 · similarity 0.86)
- Population Aware Diffusion for Time Series Generation (AAAI 2025 · similarity 0.85)
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.