PaperlayerView exact revision on arXiv

Source-linked AI summary

WaveTLM: Reliable Time-Series Language Modeling through Task Compilation

Jiahui Chen, Bingke Zhu, Hongyu Pan, Yingying Chen

arXiv:2609.18812v1cs.LG

TL;DR

Time-series language models can produce plausible responses that fail to instantiate the required task object. WaveTLM compiles requests and wave-grounded evidence into typed task states and constructs task-native outputs; on ExecTS-QA, it achieves 99.40% contract-valid coverage versus 37.83% for the strongest evaluated string-first baseline.

  • Problem

    Plausible language does not guarantee that time-series responses instantiate the required numerical tensor, legal decision, or structured record.

    Source: Introduction · p. 1

  • Method

    WaveTLM compiles requests, visible arguments, and wave-grounded evidence into typed task states, then uses task-native executors to construct outputs; ExecTS-QA evaluates five task families under deterministic contracts.

    Source: Introduction · p. 2

  • Results

    99.40% contract-valid coverage across five task families versus 37.83% for the strongest evaluated string-first baseline, while retaining competitive task-dependent utility.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    The results support task compilation as a route to reliable time-series outputs.

    Source: Conclusion · p. 7

    External results do not imply uniform predictive superiority, as performance depends on the task, dataset, and evaluation protocol.

    Source: External Transfer · p. 6

Abstract

from arXiv · show

Time-series language models provide a shared natural-language interface across temporal tasks, but plausible text does not guarantee reliable task outputs. Responses may appear reasonable while hallucinating the required object: numerical sequences can violate shape, scale, channel order, or temporal alignment, and textual decisions can fall outside the legal label space. We formulate reliable time-series language modeling, separating task-object reliability from predictive quality. We introduce ExecTS-QA, a contract-grounded benchmark spanning forecasting, imputation, classification, anomaly detection, and waveform analysis. We further propose WaveTLM, a unified compiler-executor model whose task compiler transforms user requests, visible arguments, and wave-grounded evidence into typed task states, while task-native executors construct numerical tensors, legal decisions, or structured records. On ExecTS-QA, a single WaveTLM checkpoint achieves 99.40% contract-valid coverage, compared with 37.83% for the strongest evaluated string-first baseline, while retaining balanced predictive performance across all five task families. Evaluations on SciTS, TSQA, IRTS-ToolBench, and ARFBench provide additional evidence of transfer. The code, construction scripts, and ExecTS-QA dataset will be publicly released upon publication. These results show that task compilation can convert plausible language generation into reliable time-series outputs.

Introduction

WaveTLM addresses task-object hallucination by compiling requests and temporal evidence into typed task states before constructing contract-aligned outputs. ExecTS-QA evaluates reliability across five time-series task families, separately from predictive quality.

  • Task-object hallucination occurs when plausible language fails to instantiate the required numerical tensor, legal decision, or structured record.Task-object reliability is distinct from predictive quality: an inaccurate but valid forecast remains measurable, whereas an invalid sequence does not.Source: Introduction · p. 1
  • Reliable outputs must satisfy type, shape, alignment, scale, channel order, finite-value, and decision-domain constraints that depend jointly on the request, temporal input, and task protocol.Surface controls such as JSON schemas and constrained decoding do not determine task-specific requirements such as horizon, mask alignment, inverse normalization, channel semantics, or legal labels.Source: Introduction · pp. 1–2
  • WaveTLM separates compilation from execution: its compiler binds the request and visible arguments to a contract and extracts task-relevant evidence before an executor constructs the required object.This makes the output type and construction path explicit before prediction, rather than relying on post-hoc repair.Source: Introduction · p. 2
  • ExecTS-QA pairs natural-language requests and task-native targets with deterministic contracts across five task families, validating outputs without evaluator-side repair.WaveTLM combines requests, semantic temporal patches, and multi-scale waveform evidence into typed states; task-native executors construct numerical tensors, legal decisions, and structured waveform records.Source: Introduction · p. 2
  • 99.40% contract-valid coverage on ExecTS-QA compares with 37.83% for the strongest evaluated string-first baseline, while WaveTLM retains competitive task-dependent utility.Evaluations on SciTS, TSQA, IRTS-ToolBench, and ARFBench provide additional transfer evidence.Source: Introduction · p. 2

Related Work

Prior temporal and language-model systems provide structured temporal modeling and broad language-facing task access, but generally do not infer the required native output object or isolate full-denominator contract reliability. ExecTS-QA complements existing QA benchmarks by evaluating contract-valid native outputs without evaluator-side repair.

Native temporal modeling

  • Native temporal models and foundation models construct structured outputs, but generally assume task-specific inputs; WaveTLM places native object construction behind a request-derived compiler.Examples include Salinas et al. 2020; Oreshkin et al. 2020; Gao et al. 2024; and Ansari et al. 2024.Source: Related Work · p. 2

Language models for time series

Time-series QA and reliability evaluation

Method

WaveTLM encodes semantic and raw-wave evidence, compiles visible requests into typed contract states, and uses task-native executors to construct outputs. Its contracts define valid object properties, while object reliability is assessed separately from predictive quality.

Reliable TLM Formulation

  • Contracts specify admissible object type, shape, alignment, scale, channel order, finite-value constraints, and legal decision domain; reliability requires valid native and recovered objects that remain equivalent.Reliability uses the full request denominator, while task-specific metrics separately measure predictive quality; evaluation does not repair incorrect routes or invalid objects.Source: Reliable TLM Formulation · p. 3, Task-Native Executors · p. 4

Task Compiler

  • The compiler combines language-aligned semantic patches with raw multi-scale, frequency, event-local, and channel-aware evidence, then creates a route-conditioned state specifying evidence, output space, and executor path.It does not directly predict the final object; structural metadata preserves channel, position, scale, and token type.Source: Task Compiler · p. 3

Task-Native Executors

  • Task-native executors construct numerical tensors, legal decisions, or structured waveform records, and validate and serialize outputs without modifying predictions to force contract compliance.Numerical executors preserve forecast horizon and channel organization or imputation-to-missing-position correspondence, inverting reversible normalization before serialization; waveform heads construct attributes while the language model generates explanations.Source: Task-Native Executors · pp. 3–4

Multitask Learning and Inference

  • The frozen language backbone and semantic patch encoder are paired with trainable adapters, resamplers, typed-output embeddings, task executors, and task-specific LoRA paths; inference routes only from visible inputs.No evaluator-side task label, route correction, output repair, scale correction, or label remapping is used.Source: Multitask Learning and Inference · p. 4

Experiments

WaveTLM pairs task compilation with task-native execution to produce reliable time-series objects while retaining predictive utility across five task families. Experiments assess contract validity, predictive quality, transfer, and component contributions.

Experimental Setup

  • ExecTS-QA contains 35,323 training and 4,795 evaluation instances, each governed by a deterministic task-native output contract.The benchmark spans forecasting, imputation, classification, anomaly detection, and waveform analysis; validity checks include type, shape, alignment, scale, channel order, and legal decision domain.Source: Experimental Setup · pp. 4–5

Task-Object Reliability

  • 99.40% CVC for WaveTLM exceeds ChatTS-SFT’s 37.83% under identical requests and strict contracts.WaveTLM produced 4,766 valid objects from 4,795 evaluation instances; the remaining cases were one runtime failure and 28 waveform-routing failures.Source: Table 2 · p. 5, Task-Object Reliability · p. 5

Task-Specific Predictive Utility

  • WaveTLM is the only evaluated language-facing system with measurable utility across all five task families and leads classification and waveform-category accuracy.UniTS has lower forecasting and imputation MAE with task-specific native inputs, while ChatTS-SFT has higher anomaly F1 among its valid decisions; predictive metrics are reported only on contract-valid objects.Source: Table 3 · p. 5, Task-Specific Predictive Utility · pp. 5–6

Task-Object Reliability

  • A tolerance audit found a maximum round-trip discrepancy of 4.68 × 10−7 for 77 flagged imputation objects, which were treated as evaluator-equivalent.The comparison followed the registered inverse transformation.Source: Task-Object Reliability · p. 5

Mechanism and Ablation Studies

  • Task-native construction reaches 100% strict numerical validity, unlike free or grammar-constrained decoding and structured spans, which do not ensure valid tensor structure or alignment.Task-specific executors improve forecasting and imputation MAE under a shared language path but reduce anomaly F1 to zero; task-specific LoRA paths restore classification accuracy to .706 and anomaly F1 to .847.Source: Figure 3 · p. 6, Mechanism and Ablation Studies · p. 6, Table 5 · p. 7
  • The selector routes all forecasting, imputation, classification, and anomaly requests correctly, plus 530 of 559 waveform requests, for 99.40% route accuracy.The Wave Adapter reduces forecasting and imputation MAE and improves classification, the route-conditioned resampler contributes most to anomaly detection, and a linear projector scores highest on waveform reasoning; effects are task dependent.Source: Mechanism and Ablation Studies · pp. 6–7

External Transfer

  • WaveTLM leads displayed TSQA-50 results on four tasks after supervised adaptation, tops IRTS-ToolBench overall and temporal-relation scores, and improves all reported ARFBench metrics.TSQA uses 1,000 adaptation examples per task; the unchanged central checkpoint is used for IRTS-ToolBench and ARFBench.Source: External Transfer · p. 6, Table 4 · p. 7
  • On SciTS without specific adaptation, WaveTLM achieves 100% numerical success, leads anomaly F1 on MEU01 and URU04, ranks second on PHU04, and is weaker on PHU05.Numerical success measures valid numerical response production, not prediction accuracy.Source: External Transfer · p. 6, Table 4 · p. 7

Conclusion

WaveTLM treats task-object hallucination as a reliability problem and uses task compilation to produce reliable time-series outputs. ExecTS-QA separates contract-valid coverage from predictive quality to measure this failure.

Conclusion

  • 99.40% contract-valid coverage across five task families for a single WaveTLM checkpoint versus 37.83% for the strongest string-first baseline.WaveTLM retains utility across four benchmarks and strong crossbenchmark transfer evidence.Source: Conclusion · p. 7
  • Task-object hallucination occurs when a plausible response fails to instantiate the numerical tensor, legal decision, or structured record required by its task.ExecTS-QA measures this failure by separating contract-valid coverage from predictive quality.Source: Conclusion · p. 7
  • WaveTLM compiles requests and wave-grounded evidence into typed task states, then constructs task-native outputs through task-specific paths.The conclusion presents task compilation as a route to reliable time-series outputs.Source: Conclusion · p. 7

Related papers

Loading 2609.18812v1…