![[2602.19455] SenTSR-Bench: Thinking with Injected Knowledge for Time-Series Reasoning](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.19455] SenTSR-Bench: Thinking with Injected Knowledge for Time-Series Reasoning
Summary
The paper introduces SenTSR-Bench, a framework that enhances time-series reasoning by integrating insights from specialized time-series language models into general reasoning models, achieving improved diagnostic accuracy.
Why It Matters
Time-series reasoning is crucial for various applications, yet existing models struggle with the balance between domain-specific knowledge and general reasoning capabilities. This research addresses that gap, providing a novel approach that could significantly enhance diagnostic processes in industrial applications, making it relevant for both AI development and practical implementations in the field.
Key Takeaways
- Introduces a hybrid knowledge-injection framework for time-series reasoning.
- Demonstrates significant performance improvements over existing models.
- Releases SenTSR-Bench, a benchmark for evaluating time-series diagnostic reasoning.
- Utilizes reinforcement learning for efficient knowledge extraction.
- Addresses the challenge of integrating domain-specific knowledge into general models.
Computer Science > Machine Learning arXiv:2602.19455 (cs) [Submitted on 23 Feb 2026] Title:SenTSR-Bench: Thinking with Injected Knowledge for Time-Series Reasoning Authors:Zelin He, Boran Han, Xiyuan Zhang, Shuai Zhang, Haotian Lin, Qi Zhu, Haoyang Fang, Danielle C. Maddix, Abdul Fatir Ansari, Akash Chandrayan, Abhinav Pradhan, Bernie Wang, Matthew Reimherr View a PDF of the paper titled SenTSR-Bench: Thinking with Injected Knowledge for Time-Series Reasoning, by Zelin He and 12 other authors View PDF HTML (experimental) Abstract:Time-series diagnostic reasoning is essential for many applications, yet existing solutions face a persistent gap: general reasoning large language models (GRLMs) possess strong reasoning skills but lack the domain-specific knowledge to understand complex time-series patterns. Conversely, fine-tuned time-series LLMs (TSLMs) understand these patterns but lack the capacity to generalize reasoning for more complicated questions. To bridge this gap, we propose a hybrid knowledge-injection framework that injects TSLM-generated insights directly into GRLM's reasoning trace, thereby achieving strong time-series reasoning with in-domain knowledge. As collecting data for knowledge injection fine-tuning is costly, we further leverage a reinforcement learning-based approach with verifiable rewards (RLVR) to elicit knowledge-rich traces without human supervision, then transfer such an in-domain thinking trace into GRLM for efficient knowledge injection. We fu...
