Research map / Time series / Foundation models for time series
Time Series Reasoning via Process-Verifiable Thinking Data Synthesis and Scheduling for Tailored LLM Reasoning
Jiahui Zhou, Dan Li, Boxin Li, Xiao Zhang, Erli Meng, Lin Li, Zhuomin Chen, Jian Lou, See-Kiong Ng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on TSRBench process-verifiable CoT with SFT+RL instead of base model | Qwen2.5-3B-Instruct · TSRBench scenario-based overall | accuracy (%): 40.99 -> 82.86 | Improved |
| train with VeriTime instead of base model | Qwen3-4B-Instruct · TSRBench scenario-based overall | accuracy (%): 75.48 -> 86.55 | Improved |
| use TSRBench CoT reasoning data instead of original concise ChatTS Q&A pairs for SFT (ablation) | Qwen3-4B-Instruct · TSRBench inferential calculation | accuracy (%): 72.38 -> 77.14 | Improved |
| add GRPO RL stage after SFT (ablation) | VeriTime on Qwen3-4B-Instruct · TimeSeriesExam | accuracy (%): 40.31 -> 47.27 | Improved |
| add GRPO RL stage after SFT (ablation) | VeriTime on Qwen2.5-3B-Instruct · DROP | accuracy (%): 66.44 -> 82.55 | Improved |
| apply TS-tailored CoT with VeriTime to reduce reasoning length | Qwen3-4B-Instruct · 8 tasks inference | average token count: 2192.46 -> 632.67 (anomaly detection) | Improved |
| remove hard reward (final answer correctness) from RL (ablation) | VeriTime on Qwen2.5-3B-Instruct · TSRBench inferential calculation | accuracy (%): 47.6 -> 68.6 | Improved |
| add critical pattern reward (ablation) | VeriTime on Qwen2.5-3B-Instruct · TSRBench scenario-based tasks | accuracy (%): 18.94% average drop when removed | Improved |
| add answer alignment and verification rewards (ablation) | VeriTime on Qwen2.5-3B-Instruct · TSRBench inferential calculation | accuracy (%): 23.62% drop when removed | Improved |
| add structural rewards (format and length) (ablation) | VeriTime on Qwen2.5-3B-Instruct · TSRBench scenario-based tasks | accuracy (%) | Improved |
| model-guided data scheduling (failed cases to RL, successful to SFT) instead of random split (ablation) | VeriTime on Qwen3-4B-Instruct · TSRBench scenario-based overall | accuracy (%): 83.30 -> 86.55 | Improved |
| apply full RL on all scenario-based data instead of model-guided scheduling (ablation) | VeriTime · TSRBench scenario-based tasks | accuracy (%) | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 on most shared results.
- TimeOmni-1 — TimeOmni-1: Incentivizing Complex Reasoning with Time Series in Large Language Models (ICLR 2026) · better on 1 of 1 shared result
- TimesNet — TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis (ICLR 2023) · better on 0 of 1 shared result
- Autoformer — Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (NeurIPS 2021) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Time-LLM: Time Series Forecasting by Reprogramming Large Language Models (ICLR 2024 · similarity 0.89)
- Bag of Tricks for Inference-time Computation of LLM Reasoning (NeurIPS 2025 · similarity 0.88)
- RL of Thoughts: Navigating LLM Reasoning with Inference-time Reinforcement Learning (ICLR 2026 · similarity 0.88)
- When More is Less: Understanding Chain-of-Thought Length in LLMs (ICLR 2026 · similarity 0.88)
- Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.88)
- Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.88)
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.
