Study model composition across expert towers

Investigate how different backbone choices for the perception, reasoning, and forecasting expert towers affect the performance of unified time-series and language models.

Background

TimeBraid uses separate language reasoning, time-series perception, and time-series forecasting experts, with the perception and forecasting experts able to share an underlying time-series model in a dual-tower architecture. The paper also evaluates a tri-tower variant with separate time-series backbones but observes no significant performance improvement from the additional backbone in its ablation study.

The authors explicitly leave broader model-composition choices unresolved, including the selection and arrangement of backbones across the expert towers. Resolving this could clarify whether specialized pretrained models should be assigned to particular roles in unified time-series–language systems.

References

The choice of backbone for each expert tower is not limited under this framework, and we leave the study of model composition as future work.

— TimeBraid: Unifying Time Series and Language for Understanding and Forecasting  (2609.29792 - Wang et al., 24 Sep 2026) in Section 3, subsection “Unified Time-Series and Language Modeling,” paragraph “Tri-Expert Architecture”

We leave the study of using different time-series foundation models for different experts for future work.

— TimeBraid: Unifying Time Series and Language for Understanding and Forecasting  (2609.29792 - Wang et al., 24 Sep 2026) in Section 5, subsection “Ablation Studies,” paragraph “RQ2: Expert Separation”