Scalability of latent-reasoning performance across larger language backbones

Determine whether the observed reduction in sensitivity to the sensor-token fusion pathway under Lapras continues to hold for larger language backbones, such as 7B-parameter models.

Background

Appendix C compares OpenTSLM-Flamingo and OpenTSLM-SoftPrompt, which share a 1B LLM but use different mechanisms for integrating sensor tokens. Lapras improves accuracy for both variants and substantially narrows their performance gap relative to direct answering.

The authors explicitly limit this conclusion to the evaluated 1B variants and leave unresolved whether the same behavior generalizes to larger LLMs.

References

However, whether this behavior holds for larger language backbones (e.g., 7B) remains open for future work.

— Lapras: Latent Reasoning for Time Series Language Models  (2610.11111 - Chen et al., 8 Oct 2026) in Appendix C, “Architecture Variant Discussion,” p. 15