Skip to content

BRIDGE: Bootstrapping Text to Control Time-Series Generation via Multi-Agent Iterative Optimization and Diffusion Modeling

🕒 Published (v1): 2025-03-04 09:40 UTC · Source: Arxiv · Venue: ICML 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

BRIDGE introduces a two-stage framework for text-controlled time-series generation (TSG): a role-based multi-agent iterative optimization harness that synthesizes paired text–time-series datasets, and a hybrid diffusion model conditioned on both LLM-encoded text and learned semantic prototypes. It achieves state-of-the-art generation fidelity on 11 of 12 cross-domain datasets and improves controllability by up to 12% MSE / 6% MAE over text-free baselines.

Problem

Most TSG methods operate unconditionally within a single domain and cannot satisfy instance-level constraints (e.g., generate ECGs matching a specific patient profile). A key blocker is the near-total absence of high-quality, fine-grained text–time-series paired data: existing datasets provide only coarse domain labels or high-level annotations. Rule-based description generation and simple LLM calls (e.g., GPT-4o + STL decomposition) fail to capture domain-specific nuance, leaving no scalable path to training a controllable TSG model.

Method

Stage 1 – Text-to-TS Data Preparation (multi-agent harness): 1. Text Template Collection: A single ReAct-style agent reasons over web sources (Google, Wikipedia) to iteratively collect TS-description articles, from which an LLM extracts 50 domain-agnostic templates, stripping all dataset-specific information. 2. Automatic Evaluation: Generated descriptions are assessed by prompting LSTPrompt/LLMTime (LLM-based zero-shot forecasters) with chain-of-thought reasoning; better text lowers forecasting MAE/MSE, providing a compute-efficient quality signal without model fine-tuning. 3. Feedback-driven Refinement: Two independent multi-agent teams (each containing a planner, scientist, engineer, and observer) iteratively improve templates (intra-group), then a manager moderates inter-group dialogue until consensus is reached. A post-processing step deduplicates and generalises the final template library.

Stage 2 – Text-Controlled TSG (BRIDGE diffusion model): - A Prototype Assignment Module maps target-domain time series to a learned dictionary \(P \in \mathbb{R}^{N_p \times d}\), producing domain prototype weights \(m\). - A hybrid conditioning signal fuses LLM-encoded text embeddings \(l\) with \((P, m)\) and is injected into cross-attention layers of a diffusion backbone. - Denoising objective: \(\mathcal{L} = \mathbb{E}_{x_0, \epsilon, n}\bigl[\|\epsilon - \epsilon_{\theta,P}(x_n, n, m, l)\|^2\bigr]\). - During inference on unseen domains, target-domain samples supply the prototype extraction; no domain label is needed.

Key Contributions

  • Multi-agent iterative pipeline for automatic, generalizable text–TS dataset synthesis, operating without task-specific fine-tuning.
  • Hybrid conditioning strategy (text + semantic prototypes) bridging the discrete–continuous modality gap in diffusion-based TSG.
  • Empirical demonstration that text quality directly transfers: descriptions that improve LLM-based TSF performance also improve TSG controllability.
  • Dataset-agnostic template library: trained on 2 datasets, evaluated on 12 disjoint datasets covering in-domain and out-of-domain settings.

Results

  • Fidelity: BRIDGE achieves SOTA on 11 of 12 datasets (measured by MDD and KL divergence) versus TimeVQVAE, TimeGAN, GT-GAN, TimeVAE baselines.
  • Controllability: Up to 12% improvement in MSE and 6% improvement in MAE (J-FTSD metric) over no-text-input generation.
  • Text quality impact: Multi-agent refined text achieves at least 15% lower MAE in the automatic evaluation task vs. unrefined (initial template) text.
  • Human evaluation: Outputs ranked favorably in both HE-Rank (model-only comparison) and HE-Mixed (with random distractors) settings.

Limitations

  • Prototype extraction at inference requires access to samples from the target domain's time series, limiting truly zero-shot deployment.
  • Template library fixed after refinement; domain shifts not seen during refinement may not be well covered.
  • Automatic evaluation proxy (LLM-based TSF quality) may not perfectly correlate with TSG fidelity in all domains.
  • Evaluation truncated in the provided excerpt (Section 5.2 onward is cut off); full ablation and out-of-domain numbers are not fully visible.
  • Channel-independent (univariate) formulation limits applicability to multivariate settings with inter-channel dependencies.

Relevance to Harnesses / Meta-Harnesses

The multi-agent refinement loop (template collection → LLM-evaluation → intra-group refinement → inter-group consensus → post-processing) is a concrete instance of a meta-harness: an automated system that orchestrates multiple LLM agents in structured roles to iteratively improve the output of another pipeline (here, data synthesis). The use of a compute-efficient surrogate signal (zero-shot TSF loss) as the inner evaluation criterion, rather than expensive fine-tuning at each iteration, is a reusable design principle for harnesses that must optimize prompt/template quality in data-scarce regimes. The role taxonomy (manager, planner, scientist, engineer, observer) and the two-team adversarial-then-consensus structure are directly applicable patterns for any harness that bootstraps training data or refines outputs via multi-agent debate.