ConceptTS: LLM-Guided Concept Bottlenecks for Interpretable Multivariate Time-Series Forecasting
ConceptTS turns LLM-proposed concepts into executable labels and interpretable forecasting bottlenecks, enabling concept-level inspection and intervention.
**ConceptTS** asks an LLM to propose task-relevant concepts and executable labeling rules, avoiding manual concept annotation. It organizes predictions through **3 bottlenecks** covering history, local forecast intervals, and the full horizon before a shared decoder produces the forecast.
For data agents, the useful pattern is to convert language-model domain knowledge into inspectable supervision rather than letting the LLM make the final prediction. Named activations can support debugging and direct concept-level interventions.
**ConceptTS** asks an LLM to propose task-relevant concepts and executable labeling rules, avoiding manual concept annotation. It organizes predictions through **3 bottlenecks** covering history, local forecast intervals, and the full horizon before a shared decoder produces the forecast. For data agents, the useful pattern is to convert language-model domain knowledge into inspectable supervision rather than letting the LLM make the final prediction. Named activations can support debugging and direct concept-level interventions. Evaluation is limited to the **Beijing Multi-Site Air Quality dataset**, where accuracy was described only as competitive with black-box baselines. The material does not quantify accuracy, labeling-rule errors, or transfer to other time-series domains.
ConceptTS turns LLM knowledge into an inspectable intermediate representation rather than trusting the model as the forecaster. Against prior context work, it extends the case for domain-native structure: generated concepts and executable rules become a debuggable interface to a conventional decoder. The evidence remains narrow, so rule quality and transfer—not merely concept readability—become the key validation gaps.