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

AI summary

Overview

Research area: Machine learning, specifically time-series reasoning with large language models (LLMs), combining time-series specialist models with general reasoning LLMs, plus a new evaluation benchmark for industrial diagnostic reasoning.

Technical level: Intermediate. Readers should be comfortable with the general ideas of LLM fine-tuning, reinforcement learning with verifiable rewards (RLVR), and benchmark construction, though the paper explains its framework conceptually as well as formally.

Scope (one sentence): The paper proposes a "knowledge injection" framework that inserts a time-series specialist model's analysis into a general reasoning model's thinking trace, trains the specialist with reinforcement learning so no human-annotated reasoning traces are needed, and releases SenTSR-Bench, a real-world multivariate sensor benchmark for diagnostic reasoning.

What This Paper Is About

Time-series diagnostic reasoning — figuring out what happened in a sensor stream, why it happened, and what to do about it — is a core capability for applications like machine monitoring. Today's tools fall into two unsatisfying camps: general reasoning LLMs reason well but lack the domain knowledge to read complex sensor patterns, while fine-tuned time-series LLMs (TSLMs) understand the patterns but overfit to narrow templates and generalize poorly. The paper's goal is to combine the two by injecting the specialist's in-domain knowledge directly into the general reasoner's reasoning trace, and to evaluate this on a new, human-annotated benchmark built from real industrial sensor data.

Key Contributions

1. A new paradigm for time-series reasoning. The authors formalize a framework in which knowledge from a time-series language model (denoted π^T) is injected into the reasoning process of a general reasoning language model (denoted π^G), steering the reasoning trace with domain knowledge without updating the reasoner's weights (the GRLM is "frozen").

2. An RL-based method for efficient knowledge injection. Rather than collecting expensive supervision for injection fine-tuning, the authors use reinforcement learning with verifiable rewards to elicit analysis-first thinking traces from the TSLM without human supervision, then transfer those traces into the GRLM at inference.

3. A real-world benchmark. They release SenTSR-Bench (Sensor-based Time-Series Diagnostic Reasoning Benchmark), described as a first-of-its-kind collection of de-identified, real-world multivariate sensor streams paired with human-annotated diagnostic text and multi-stage questions.

4. Broad empirical validation. Evaluations on SenTSR-Bench and two public benchmarks (TSEvol and the TS&Language Benchmark's MCQ2) show gains over stand-alone specialists and reasoners across open-source and closed-source models.

Main Findings

• Injection beats both stand-alone baselines. On SenTSR-Bench, knowledge injection improves over the specialized TSLM by +15.5% to +26.1% and over the zero-shot general reasoner (GRLM) by +7.3% to +22.4%. Improvements span all three tasks ("What happened", "How happened", "Suggested fix") and are largest on "How happened", which the authors attribute to needing both in-domain anomaly knowledge and strong causal reasoning.

• Public benchmarks show the same direction. On TSEvol and TS&Language (MCQ2), injection yields +5.2% to +10.4% over the specialist and +2.7% to +10.4% over the reasoner.

• RL-based injection produces consistently larger gains than SFT-based injection. Measured against zero-shot GRLMs, the RL-based "thinking transfer" approach delivers 1.66x the improvement on Claude3.7-Vision, 2.00x on Qwen3-32B, and 2.92x on DeepSeekR1-Distilled-Qwen-32B. It is also reported as 3.27x better than few-shot prompting on Claude3.7-Vision.

• Headline numbers on SenTSR-Bench. With Claude3.7-Vision, RL-Injection reaches an overall score of 0.695 ± 0.012 versus 0.620 ± 0.006 for zero-shot Claude3.7-Vision and 0.551 ± 0.014 for the RL-trained Qwen-VL-3B TSLM. With Claude3.7-Text, RL-Injection reaches 0.650 ± 0.010 versus 0.531 ± 0.011 zero-shot and 0.549 ± 0.019 for the SFT TSLM.

• Early injection is the strongest strategy. Comparing early, intermediate, and late injection using ChatTS-14B as the specialist, early injection gave the best results on both reasoners tested: overall 0.736 ± 0.002 with Claude3.7-Vision and 0.715 ± 0.003 with Claude3.7-Text. The authors explain that mid/late injection requires the specialist to revise long reasoning traces, which lies outside its QA-style training distribution and causes drift or hallucination.

• Token efficiency. Tokenized multivariate time series in TSEvol can exceed roughly 50k tokens, making few-shot prompting infeasible; injection instead supplies a compact analysis snippet through thinking prefill.

• Injection beats prompting-based collaboration. Across model families and all three benchmark settings, placing knowledge inside the reasoning trace outperformed providing the same TSLM output as external prompt instructions.

• Robustness when the specialist is weak. Even when the TSLM performs poorly (the paper highlights the Inductive task), the injected model leverages the reasoner's critical thinking capacity to stay competitive.

Methodology in Plain English

The framework has two stages.

Stage I — train the specialist with reinforcement learning. A small time-series LLM (primarily Qwen2.5-VL-3B) is trained using GRPO-style reinforcement learning with verifiable rewards. Each sampled completion must follow a fixed structure — a thinking block containing a reasoning trace, followed by an answer block. Two binary rewards are used: a format reward (1 if the output follows the target structure, 0 otherwise) and a hard reward (1 if the predicted answer matches ground truth, 0 otherwise), summed into a composite reward. The key point is that no labeled reasoning traces are required — the analysis-first thinking behavior is elicited purely from structural and correctness feedback.

Stage II — inject at inference. At test time, the specialist generates a thinking trace for the question and time series. That trace, followed by a short reflection cue (for example, "Wait, let me reflect on my previous thinking process with the time-series data."), is inserted directly after the general reasoner's opening thinking token. The general reasoner then continues generating its own reasoning and final answer, conditioned on the injected knowledge. The injected text can be fed as assistant prefill for models that support it; for models that do not, the authors wrap the injected trace in the model's recommended thinking template as an instructional proxy.

The paper also defines two alternative injection points — intermediate injection (inserting specialist knowledge at low-confidence positions in the reasoner's trace) and late injection (appending a specialist critique of the full trace before the final answer) — but adopts early injection as the default.

Benchmark construction. For SenTSR-Bench, the authors filtered 110 multivariate sensor streams from an initial pool of over 2,000 candidate samples, standardized all signals to remove sensitive information, built a de-identifying annotation pipeline, and constructed 330 multiple-choice questions pairing ground-truth answers with distractors. Because few real-world seeds exist, they also built a two-stage vision-language-model pipeline to synthesize training data: Stage 1 used plots and context from 23 de-identified seeds to generate Python code mimicking the original behaviors; Stage 2 turned those simulators into compact stochastic generators with randomized dynamics and parameter variation. The synthetic data yielded 6,000 multiple-choice training entries matching the evaluation design.

Evaluation setup. General reasoners tested include the open-source DeepSeekR1-Distilled-Qwen-32B and Qwen3-32B, and closed-source Claude3.7 with time series encoded either as vision (-vision) or text (-text). Fine-tuned specialists used were primarily Qwen2.5-VL-3B, with ChatTS-14B used for injection-design exploration. Generative QA tasks (inductive reasoning in SenTSR-Bench) were evaluated with RAGAS; verifiable tasks report accuracy; all results are averaged over three independent runs.

Why This Matters

Impact on research. The paper reframes time-series reasoning as a collaboration problem rather than a single-model problem: instead of trying to make one model both a domain expert and a general reasoner, it routes expertise from a specialist into a reasoner's thinking trace. It also shows that verifiable-reward RL can supply the "middle" supervision (reasoning traces) that diagnostic datasets normally lack, since such datasets typically contain only ground-truth answers. The release of a human-annotated, real-world multivariate sensor benchmark addresses a gap the authors document in their comparison with TSEvol, TS&Language, and MTBench, which they characterize as LLM-annotated or synthetic and single-question.

Real-world applications:

• Industrial machine monitoring: diagnosing equipment failures from vibration (acceleration, velocity) and temperature sensor streams.
• Root-cause diagnosis: moving past anomaly flagging to explain why a fault occurred.
• Maintenance action recommendation: proposing corrective actions, corresponding to the benchmark's "Suggested fix" stage.
• Diagnostic question answering over multi-stage sensor events, including characterizing which segments of a multivariate series are anomalous.

Industry relevance. The framework requires no updates to the general reasoner's weights, works with standard LLM APIs, and is compatible with both open-source and closed-source models. The authors emphasize token efficiency: injection passes a compact analysis snippet rather than the ~50k-plus tokens that tokenizing TSEvol's multivariate series can require, making it more practical than few-shot prompting for long sensor inputs.

Future Directions

• Scaling real-world seed data. The training pipeline was built from only 23 de-identified seeds, and the evaluation benchmark from 110 streams filtered out of over 2,000 candidates. The authors measure sensitivity of TSLM performance to synthetic data diversity, but the underlying scarcity of real-world seeds remains an open constraint.
• Reducing the information bottleneck. The authors ablate removing the raw series from the GRLM input and report that relying solely on the specialist's textual summary creates an information bottleneck that limits downstream reasoning, pointing to better ways of preserving raw-signal access.
• Choosing the right RL objective. GRPO is compared against DAPO, GSPO, and CISPO to test whether the gains come from injection specifically or could be replicated by stronger RL objectives alone.
• Generalizing the approach. The conclusion explicitly encourages further investigation of knowledge injection for broader time-series diagnostic reasoning tasks, beyond the industrial diagnostic setting studied here. The content provided does not include a formal limitations section; it also does not report human/expert evaluation of the generated diagnostic explanations or deployment results.

Target Audience

Researchers and practitioners working at the intersection of time series and LLM reasoning will get the most from this paper: LLM post-training and reinforcement-learning researchers interested in verifiable-reward training; time-series and anomaly-detection researchers who need interpretable, causal-style diagnoses rather than anomaly scores; benchmark developers looking for a human-annotated, real-world alternative to LLM-generated time-series QA data; and industrial machine-monitoring or predictive-maintenance teams who want a practical way to add a general reasoning model on top of an existing sensor specialist without retraining the large model.

Authors’ 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 further release SenTSR-Bench, a multivariate time-series-based diagnostic reasoning benchmark collected from real-world industrial operations. Across SenTSR-Bench and other public datasets, our method consistently surpasses TSLMs by 9.1%-26.1% and GRLMs by 7.9%-22.4%, delivering robust, context-aware time-series diagnostic insights.