Research map / Time series / Foundation models for time series

TSQAgent: Rating Time Series Data Quality via Dedicated Agentic Reasoning

NeurIPS 2026 · Time series › Foundation models for time series · approach: Time-series language reasoning · 0 citations (OpenAlex)

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.

Builds TSQBench to test LLMs on time-series quality dimension identification and pairwise quality comparison, then proposes TSQAgent (Perceiver/Inspector/Adjudicator agents with GRPO-trained dimension selection and tool-augmented analysis) that improves LLM judgments and quality-aware data selection.

Problem

LLMs' ability to adaptively identify relevant time-series quality dimensions and perform evidence-grounded quantitative quality comparisons remains unclear, and existing LLM-based TS quality rating relies on predefined dimensions and text-only reasoning.

Approach

A multi-agent framework where a GRPO-trained Perceiver selects relevant quality dimensions (precision-weighted reward), an Inspector runs bounded ReAct loops with 16 time-series analytical tools per dimension, and an Adjudicator reflects and aggregates into a pairwise judgment, which is converted to scalar scores via Bradley-Terry for data selection.

Key results

• On TSQBench, tool augmentation raises quality-comparison accuracy from about 55%-61% to 78%-84%.
• GRPO-trained dimension selection raises dimension-identification precision from about 33%-38% to nearly 60% across open-source models.
• On cross-dataset Timer-S1 (8.3B) fine-tuning, 75% of data selected by the method gives average RMSE 0.341 vs 0.338 for full data and 0.350 for random 75%.

Surprising findings

• Even the proprietary models reach only 54.7%-61.0% quality comparison accuracy, barely above random guessing.

Limitations

• Benchmark is fully synthetic, so transfer of quality definitions to real-world defects is not directly validated.
• Tool-augmented gains are reported on a synthetic benchmark; real-data gains are measured only through downstream data selection.

Experimental setup

Datasets: Electricity ExchangeRate Traffic Weather M4 MedicalImages CBF BME Handwriting

Benchmarks: TSQBench Electricity ExchangeRate Traffic Weather M4 MedicalImages CBF BME Handwriting

Metrics: precision recall accuracy RMSE MAPE

Baselines: Random DataOob DataShapley TimeInf TSRating

Builds on: GRPO ReAct TSRating TimeInf

Closest papers by meaning

• CLUES: Collaborative Private-domain High-quality Data Selection for LLMs via Training Dynamics (NeurIPS 2024 · similarity 0.85)
• TSGM: A Flexible Framework for Generative Modeling of Synthetic Time Series (NeurIPS 2024 · similarity 0.84)
• Can LLMs Understand Time Series Anomalies? (ICLR 2025 · similarity 0.84)
• InfiAgent-DABench: Evaluating Agents on Data Analysis Tasks (ICML 2024 · similarity 0.84)
• BRIDGE: Bootstrapping Text to Control Time-Series Generation via Multi-Agent Iterative Optimization and Diffusion Modeling (ICML 2025 · similarity 0.84)
• TEST: Text Prototype Aligned Embedding to Activate LLM's Ability for Time Series (ICLR 2024 · similarity 0.84)