Lapras: Latent Reasoning for Time Series Language Models

Yuliang Chen, Yu Yvonne Wu, Arvind Pillai, Nicholas C. Jacobson, Andrew Campbell

Summary

Post-training framework for time-series language models (TSLMs): instead of verbalising the signal in an explicit chain-of-thought, the model reasons through continuous "thoughts" in the joint time-series–language space and emits text only for the answer. Trained by teacher–student self-distillation: a CoT-SFT teacher reasons in text, and the student aligns its answer-stage hidden states to the teacher's. Evaluated on four TSLM backbones (ChatTS-7B, OpenTSLM-1B Flamingo variant, SLIP-1B, ITFormer-0.5B) and five QA tasks: ECG (ECG-Reasoning benchmark, Oh et al.

Reported benchmarks