PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2602.19455v1cs.LGcs.AIcs.CLstat.ML

TL;DR

Time-series diagnostic reasoning needs both domain knowledge and general reasoning, but GRLMs and TSLMs each lack one of these capabilities. The paper injects TSLM analysis into GRLM reasoning, uses RLVR to produce knowledge-rich traces, and releases SenTSR-Bench. Across multiple datasets, the method surpasses TSLMs by 9.1–26.1% and GRLMs by 7.9–22.4%.

  • Problem

    Existing approaches lack a combination of temporal and causal diagnostic reasoning, domain knowledge for complex time-series patterns, and sufficient reasoning generalization.

  • Method

    The framework injects TSLM-generated in-domain analysis into a frozen GRLM’s reasoning trace and uses RLVR to elicit analysis-first traces without human supervision.

  • Results

    9.1–26.1% over TSLMs and 7.9–22.4% over GRLMs are the reported gains across SenTSR-Bench and other benchmark datasets.

  • Takeaways & Limitations

    SenTSR-Bench enables evaluation of multivariate time-series diagnostic reasoning on real-world data with multi-stage questions and human annotations.

Abstract

from arXiv · show

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.

1 Introduction

Time-series diagnostic reasoning requires temporal and causal understanding, but GRLMs lack domain knowledge while TSLMs lack broader reasoning capacity. The proposed framework injects TSLM knowledge into GRLM reasoning, uses RLVR to elicit thinking traces, and evaluates the approach on the real-world SenTSR-Bench benchmark.

  • Motivation: Existing anomaly-detection methods identify irregularities but do not explain origins, diagnose root causes, or recommend corrective actions.The limitation arises from lacking temporal and causal reasoning.
  • Motivation: GRLMs provide reasoning ability without sufficient knowledge of complex time-series patterns, while TSLMs capture patterns but lack reasoning depth or generalization capacity.Both standalone model types therefore fall short in practice.
  • Proposed framework: The framework injects in-domain TSLM knowledge directly into a GRLM’s reasoning process, combining domain knowledge with general reasoning.When injected knowledge is weaker, the GRLM can correct it using its critical-thinking capacity.
  • Proposed framework: RLVR elicits knowledge-rich, analysis-first TSLM thinking traces without manual supervision for injection into GRLM reasoning.The approach addresses the costly need for large, high-quality knowledge-injection datasets.
  • Benchmark: SenTSR-Bench provides real-world multivariate sensor streams and diagnostic texts for evaluating time-series diagnostic reasoning.The benchmark is built from real-world diagnostic events with human-annotated data.
  • Results: 9.1–26.1% and 7.9–22.4% are the reported margins by which the method surpasses TSLMs and GRLMs, respectively, across SenTSR-Bench and other datasets.The evaluation covers both closed-source and open-source reasoning models.

2 Methodology

The methodology injects specialist time-series knowledge into a general reasoning model’s trace, using RLVR-trained thinking traces to reduce supervision requirements and support grounded diagnostic reasoning.

  • 2.1 Preliminaries and Notation: The framework represents time-series inputs and textual tasks in token space, then generates an internal reasoning trace before producing the final answer.The reasoning–then–response decomposition exposes the trace for later inspection and modification.
  • 2.2 General Knowledge Injection Paradigm: A frozen GRLM supplies general reasoning while a fine-tuned TSLM supplies in-domain time-series knowledge for injection.The TSLM is a small domain specialist, whereas the GRLM follows the reasoning–then–response factorization.
  • 2.2 General Knowledge Injection Paradigm: The TSLM is queried with an injection-oriented prompt, and its knowledge output is inserted into the GRLM’s current reasoning prefix before generation resumes.The deterministic query-shaping and injection functions determine how specialist knowledge enters the reasoning trace.
  • 2.3 Instantiating Knowledge Injection: Early injection places a help-oriented specialist trace immediately after the thinking marker, while intermediate and late variants inject corrections or critiques later.The authors find early injection broadly effective and use it as the default.
  • 2.4 Knowledge Injection with RL-Honed Thinking Traces: Thinking transfer trains the TSLM to produce analysis before an answer, aligning its training role with its deployment as an injected knowledge source.This addresses the mismatch between direct-answer training and the need for intermediate evidence during injection.
  • 2.4 Knowledge Injection with RL-Honed Thinking Traces: RLVR trains the specialist without intermediate-trace supervision by sampling completions and optimizing format and task rewards.The composite reward includes a binary format reward and a hard reward component.

3 Benchmark: SenTSR-Bench

SenTSR-Bench is a real-world, multivariate sensor benchmark designed to evaluate diagnostic reasoning beyond anomaly detection, including events, causes, and corrective actions.

  • Motivation: SenTSR-Bench addresses limited availability of high-quality datasets pairing real-world time series with textual diagnostic annotations.Existing benchmarks are described as mainly LLM-annotated public datasets or fully synthetic pairs, often with one question per series.
  • Dataset Characteristics: The benchmark uses de-identified multivariate sensor signals from vibration and temperature monitoring, paired with human-curated diagnostic annotations.The data are motivated by machine breakdown diagnosis and troubleshooting.
  • Diagnostic Tasks: Its questions cover what happened, how it happened, and suggested fixes, testing anomalous-segment recognition, root-cause inference, and corrective actions.This moves evaluation beyond anomaly flagging toward full diagnostic reasoning.
  • Evaluation Dataset Curation: Evaluation curation filters 110 multivariate sensor streams from more than 2,000 candidate samples and standardizes the signals to remove sensitive information.The pipeline selects anomalous patterns tied to potential troubleshooting actions before annotation.
  • Training Dataset Generation: Training data are diversified from 23 de-identified real-world seeds through VLM-generated simulators and randomized latent-state dynamics.The process aims to produce varied realistic series while retaining statistical and structural characteristics.

4 Experiments

Experiments evaluate knowledge injection across real-world and public benchmarks, model families, training paradigms, and injection strategies. Injection consistently improves over specialist and general-reasoning baselines, with RL-based and early injection showing particular advantages.

  • Experiment Setup: SenTSR-Bench evaluates anomaly characterization, root-cause diagnosis, and corrective-action recommendation, while TSEvol and TS&Language provide public reasoning benchmarks.
  • Performance Analysis: Across SenTSR-Bench and public benchmarks, injection boosts accuracy over both standalone specialists and standalone reasoners.
  • Performance Analysis: +5.2% to +10.4% gains over specialists and +2.7% to +10.4% over reasoners appear on the public benchmarks.
  • Performance Analysis: RL-based injection yields 1.66×–2.92× larger gains than SFT-based injection and provides a compact alternative when time-series prompts exceed ∼50k tokens.
  • Framework Analysis: Early injection yields the strongest gains because short specialist analyses align better with the specialist’s training than revising long reasoning traces.
  • Framework Analysis: Across zero-shot reasoning, prompting, and injection, injection consistently outperforms prompting across model families and benchmark datasets.

5 Related Work

Related work follows two main directions: prompting-based structured reasoning and specialist models trained on time-series–text pairs. The former lacks domain-specific priors, while the latter can overfit in-domain data and lack broader reasoning capacity.

  • Prior work studies prompting-based structured reasoning over temporal data and specialist models post-trained on time-series–text pairs.
  • Prompting-based approaches lack domain-specific priors for capturing key diagnostic patterns, whereas specialist models often overfit to in-domain data.

6 Conclusion

The paper combines time-series specialists’ domain knowledge with general LLM reasoning, adds RL-based thinking transfer, and releases a real-world diagnostic benchmark. Across SenTSR-Bench and public datasets, the framework improves over standalone baselines.

  • The knowledge injection framework combines domain knowledge from time-series specialists with the reasoning ability of general LLMs.
  • RL-based thinking transfer elicits analysis-first traces without supervision for effective, task-aligned injection.
  • SenTSR-Bench is a real-world benchmark with multi-stage questions covering anomaly recognition, root-cause diagnosis, and corrective suggestions.
  • 7.9%–26.1% improvements over standalone baselines are reported across SenTSR-Bench and public datasets.

A.1 Details of GRPO Training Objective

The appendix describes GRPO training and alternative injection procedures. GRPO normalizes group rewards into token-shared advantages and optimizes a clipped, KL-regularized objective, while intermediate and late injection target uncertain or completed reasoning traces.

  • Details of GRPO Training Objective: GRPO samples groups of complete reasoning-and-answer sequences and assigns each sequence a scalar composite reward.
  • Details of GRPO Training Objective: Rewards are normalized by subtracting the group mean and dividing by the standard deviation, with γ ensuring numerical stability.
  • Details of GRPO Training Objective: Every token in a sampled sequence shares that sequence’s normalized advantage, producing stable gradient updates across contexts.
  • Details of GRPO Training Objective: The training objective improves high-reward completions, clips updates for stability, and penalizes divergence from a frozen reference model through KL regularization.
  • Intermediate Knowledge Injection: Intermediate injection identifies the lowest-confidence sentence, asks the specialist to verify it against the series, and resumes generation after inserting evidence or correction.
  • Late Knowledge Injection: Late injection appends a specialist critique of the complete draft, after which the reasoner reconciles disagreements and updates its conclusion.

B.1 Ablation on Reliance on TSLM Textual Summaries

The ablation tests whether GRLMs should rely only on TSLM summaries or also access raw time series. Full injection combines both inputs and supports more robust diagnostic reasoning than summary-only injection.

  • Input Design: The full design gives both the TSLM and GRLM access to raw time series X, while the injected summary provides auxiliary guidance.The GRLM conditions on (X, q, r), rather than using the textual summary as its only information source.
  • Ablation Results: Summary-only injection improves over the standalone TSLM but underperforms the full injection setting.The ablation directly compares knowledge transfer without raw-series access against the proposed dual-input design.
  • Ablation Results: Approximately 7% improvement is obtained when the GRLM relies only on the TSLM-generated textual summary.This average gain is reported for the summary-only variant across the evaluated setting.
  • Interpretation: The dual-input design lets the GRLM combine its own time-series perception with injected domain knowledge instead of blindly inheriting TSLM errors.This combination is described as leading to more robust and accurate diagnostic reasoning.
  • Prompting Baselines: Structured prompting and sampling-based alternatives consistently improve over the zero-shot GRLM baseline.The comparison includes few-shot prompting, self-consistency with three runs, and tree-of-thought with three branches.

B.3 Sensitivity of TSLM performance to synthetic data diversity

The ablation examines how synthetic-data diversity affects TSLM performance and compares reinforcement-learning objectives used for TSLM training. Training is essential, but performance largely saturates before the full synthetic dataset.

  • Synthetic Data Diversity: Without synthetic data, TSLM performance remains close to random guessing across all subtasks.The result indicates that training is needed to establish basic time-series diagnostic reasoning ability for the small model.
  • Synthetic Data Diversity: Roughly 50% of seed-generated synthetic data recovers the majority of final performance.Increasing the synthetic-data proportion beyond this point produces progressively smaller gains.
  • Synthetic Data Diversity: Beyond 75% synthetic-data diversity, additional performance gains are marginal across subtasks.Saturation is slightly stronger for higher-level reasoning tasks, including How Happened and Suggested Fix.
  • RL Objectives: The study compares GRPO, DAPO, GSPO, and CISPO as reinforcement-learning objectives for TSLM training.The comparison evaluates whether more advanced objectives improve training in this setting.
  • RL Objectives: DAPO yields faster and smoother reward convergence than the GRPO baseline.The supplied passage introduces this reward-trajectory comparison but does not state the final-reward outcome.

D.2.1 Evaluation Dataset Curation

The benchmark curation pipeline builds a privacy-preserving diagnostic dataset from real industrial sensor streams. It combines anomaly selection, human annotation, and structured multi-stage queries with synthetic-data generation for training.

  • Signal Selection and Preprocessing: Approximately 2,000 multivariate sensor streams are screened, yielding 110 streams with clear anomalous behaviors tied to troubleshooting events.Selected anomalies include persistent deviations, sharp drops or spikes, and sudden periodicity shifts.
  • Human Annotation Pipeline: Human experts annotate anomalous windows with observed patterns, plausible root causes, and candidate corrective actions.Annotators receive sanitized time-series segments and high-level machine categories to protect proprietary context.
  • Evaluation Query Construction: Evaluation queries follow three stages: identifying the anomalous segment, inferring its root cause, and suggesting a corrective action.Distractors are sampled from other anomaly clusters, requiring recognition and reasoning rather than memorization.
  • Benchmark Outcome: The resulting benchmark uses human-authored annotations grounded in real sensor signals while emphasizing diagnostic depth and privacy protection.This multi-stage curation produces a realistic and challenging evaluation setting.
  • Training Data Generation: A VLM-based two-stage pipeline bootstraps realistic simulators from a small set of seed signals for scalable training-data construction.The pipeline addresses scarce and complex real-world multivariate sensor data.

E.1 Implementation Details: Reasoning Model Baselines

The implementation evaluates reasoning baselines with standardized time-series encodings, long-context adaptation, and hosted model infrastructure. It also tests multiple positions and mechanisms for injecting TSLM knowledge into GRLM reasoning.

  • Reasoning Baselines: Baselines are evaluated under zero-shot and few-shot prompting, with three randomly sampled SenTSR-Bench demonstrations in few-shot experiments.Models are served through an OpenAI-compatible vLLM server using HuggingFace checkpoints.
  • Time-Series Encoding: Multivariate time series are encoded either as stacked line plots or structured JSON-like channel sequences.The textual representation preserves value alignment across channels and available column names.
  • Long-Context Adaptation: RoPE scaling enables open-source models to process certain multivariate inputs exceeding 50k tokens without truncation.The adaptation is intended to preserve grounding in long time-series signals.
  • TSLM Training: TSLMs are initialized from Qwen-VL-3B-Instruct and trained with supervised fine-tuning followed by reinforcement learning with verifiable rewards.GRPO is used to elicit analysis-first completions without explicit thinking supervision.
  • Knowledge Injection: For closed-source models, injected traces are wrapped in thinking-template tags and followed by instructions to continue the reasoning process.The proxy is used when editable internal thinking buffers are unavailable.
  • Knowledge Injection: Early injection inserts TSLM-generated structured observations at the start of the GRLM reasoning process.Intermediate injection requests TSLM clarification during low-confidence reasoning, while late injection reviews and revises a completed trace.

F Additional Case Study

The case studies show that knowledge injection combines TSLM signal interpretation with GRLM reasoning to correct diagnostic errors. Integrating domain analysis directly into the reasoning flow is more effective than providing it only as an external prompt.

  • Knowledge injection corrected both standalone models’ misreadings and produced the correct final diagnosis.The TSLM hallucinated a joint signal increase, while the GRLM assumed a late temperature rise; injection preserved accurate observations and corrected the reasoning flaw.
  • The injection-based approach integrated TSLM analysis directly into the GRLM’s reasoning flow, enabling joint exploration and progressive hypothesis narrowing.
  • The qualitative studies compare standalone TSLM, standalone GRLM, knowledge prompting, and knowledge injection on diagnostic reasoning examples.
  • External knowledge prompting led the GRLM to reason largely in isolation and use the TSLM analysis mainly for end-stage validation.
Loading 2602.19455v1…