💬 NLP

SenTSR-Bench: Thinking with Injected Knowledge for Time-Series Reasoning

This paper introduces SenTSR-Bench, a real-world industrial time-series diagnostic reasoning benchmark, and proposes a hybrid framework that leverages reinforcement learning to inject domain-specific insights from fine-tuned time-series models into general reasoning LLMs, significantly outperforming existing models in complex diagnostic tasks.

Original 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

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original 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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Smart Generalist" vs. The "Specialist"

Imagine you are trying to diagnose why a massive, complex machine in a factory is making a weird noise and shaking. You have two experts to help you:

The Generalist Detective (GRLM): This is a super-smart AI (like a brilliant human detective) who is great at logic, solving puzzles, and connecting dots. However, they have never seen a factory machine before. They know how to think, but they don't know what a "vibration spike" actually means in a machine. If you show them a graph, they might guess the wrong cause because they lack specific factory knowledge.
The Factory Specialist (TSLM): This is an AI that has studied thousands of factory machines. They know exactly what a broken gear looks like on a graph. However, they are bad at explaining their reasoning. They might say, "It's broken," but they can't explain why or how to fix it logically. They also tend to get confused if the machine is slightly different from what they've seen before.

The Gap: The Detective is too smart but clueless about the machine. The Specialist knows the machine but can't think through the complex "why" and "how" questions.

The Solution: "Knowledge Injection"

The authors of this paper came up with a clever way to combine these two experts. They call it Knowledge Injection.

Instead of asking the Detective to guess, or asking the Specialist to just give an answer, they let the Specialist whisper the facts directly into the Detective's ear while the Detective is thinking.

How it works: The Detective starts thinking, "Hmm, the machine is shaking..."
The Injection: Before the Detective finishes that thought, the Specialist jumps in and says, "Wait! That shaking pattern usually means a loose belt, not a broken engine."
The Result: The Detective takes that specific fact, uses their super-smart logic to figure out the root cause, and suggests the perfect fix.

The Detective gets the facts from the Specialist and the reasoning from their own brain.

The Secret Sauce: "Thinking Transfer" via Reinforcement Learning

There was one big problem with this plan. The Specialist AI was trained to just give a final answer ("It's a broken belt"). It wasn't trained to explain its thinking process ("I see a spike here, which means..."). If you just asked it for "thinking," it would make things up (hallucinate) because it wasn't used to doing that.

To fix this, the authors used a technique called Reinforcement Learning with Verifiable Rewards (RLVR).

The Analogy: Imagine training a dog. Usually, you only give it a treat when it sits. But here, they wanted the dog to "think out loud" before sitting.
The Trick: They didn't need a human to write down the dog's thoughts. Instead, they set up a game where the AI gets a "reward" (points) if:
1. It follows a specific format (it actually writes down its thoughts).
2. The final answer it gives is correct.
The Result: The AI learned, on its own, to generate high-quality "thinking traces" (explanations) just to get the points. They then took these "thinking traces" and injected them into the Detective. This is called Thinking Transfer.

The New Playground: SenTSR-Bench

To prove this works, the team couldn't just use fake data. They needed a real-world test. They built SenTSR-Bench.

What is it? A giant test bank of real machine data (vibration, temperature) from actual factories, paired with human-written diagnostic stories.
Why is it special? Previous tests were either made up by computers or only asked simple questions like "Is this broken?" SenTSR-Bench asks the hard, multi-step questions:
1. What happened? (Spot the weird pattern).
2. How did it happen? (Find the root cause).
3. What should we do? (Suggest a fix).

The Results

When they tested their new system (The Detective + The Specialist + The Injection):

It beat the Specialist by a huge margin (up to 26% better) because the Detective could now reason through the Specialist's facts.
It beat the Detective by a huge margin (up to 22% better) because the Detective finally had the right facts to work with.
It was much better than just asking the Detective to "read a report" from the Specialist (which is like the old way of prompting). Injecting the knowledge inside the thinking process worked much better.

Summary

Think of it like a medical diagnosis:

The Generalist is a brilliant medical student who knows how to diagnose diseases but has never seen a specific rare virus.
The Specialist is a lab technician who can identify the virus instantly but can't write a treatment plan.
The Paper's Method: The technician hands the student a note while the student is writing the diagnosis, saying, "This virus causes X symptom." The student then uses their medical knowledge to write the perfect treatment plan.

The paper shows that by combining the logic of big AI with the specific knowledge of small, specialized AI, we can solve complex time-series problems that neither could solve alone.

...` format.
* Hard Reward ( $r_{hard}$ ): Checks if the final answer matches the ground truth.

Outcome: The TSLM learns to generate "analysis-first" traces without explicit human-labeled reasoning data. These RL-honed traces are then injected into the GRLM at inference time.

3. Key Contributions

New Paradigm for Time-Series Reasoning:
Formalized a framework that injects in-domain knowledge from a specialist (TSLM) into the reasoning process of a general model (GRLM), steering the reasoning trajectory with domain facts while preserving generalization.
RL-Based Efficient Injection:
Proposed a method to elicit knowledge-rich thinking traces from TSLMs using Reinforcement Learning with Verifiable Rewards (RLVR). This eliminates the need for costly, manually annotated reasoning datasets, enabling scalable "thinking transfer."
SenTSR-Bench Benchmark:
Released SenTSR-Bench, the first real-world, multivariate time-series diagnostic reasoning benchmark.
- Source: Derived from de-identified industrial sensor data (vibration, temperature).
- Structure: Features multi-stage diagnostic questions: (1) What happened? (Anomaly recognition), (2) How happened? (Root-cause diagnosis), and (3) Suggested fix? (Corrective action).
- Quality: Human-verified annotations, contrasting with prior benchmarks that rely on LLM-generated or synthetic data.

4. Experimental Results

The method was evaluated on SenTSR-Bench and public datasets (TSEvol, TS&Language) using various models (e.g., Claude 3.7, Qwen, DeepSeek R1).

Performance Gains:
- The proposed method consistently outperforms standalone TSLMs by 9.1% – 26.1%.
- It outperforms standalone GRLMs by 7.9% – 22.4%.
- RL-based Injection yields significantly larger gains (1.66× to 2.92×) compared to Supervised Fine-Tuning (SFT) based injection, demonstrating the efficacy of the RL training approach.
Comparison with Prompting:
- Knowledge injection significantly outperforms few-shot prompting and prompt-based collaboration. Injection places knowledge inside the reasoning flow, forcing the model to interact with it, whereas prompting often leads to the model ignoring external context.
Ablation Studies:
- Removing direct time-series access from the GRLM (relying only on TSLM summaries) causes a performance drop, confirming that the GRLM needs raw data to validate injected knowledge.
- Early Injection is identified as the most robust strategy, as mid/late injection requires the specialist to process long reasoning traces, which can lead to hallucinations.

5. Significance

Bridging the Gap: The work successfully bridges the divide between specialized domain models and general reasoning models, offering a practical solution for complex industrial diagnostics where neither model type suffices alone.
Scalability: By using RLVR to generate training data for the injection process, the framework reduces reliance on expensive human annotation, making it scalable for diverse time-series domains.
Benchmarking: SenTSR-Bench sets a new standard for evaluating time-series reasoning, moving beyond simple anomaly detection to complex, multi-stage diagnostic reasoning with human-verified ground truth.
Practical Impact: The approach delivers robust, context-aware diagnostic insights, directly applicable to predictive maintenance, healthcare monitoring, and industrial IoT operations.