Lapras (latent-reasoning post-training for time-series language models)
ECGIMUsleep-signaltime-seriesTextOpen Weights
Key facts
Summary
Post-training recipe that gives existing TSLMs (ChatTS-7B, OpenTSLM-1B, SLIP-1B, ITFormer-0.5B) latent ("continuous-thought") reasoning distilled from a CoT teacher.
Reported benchmarks
Human activity recognition (IMU)
opentslm-har-cot · Macro-F1 ↑
#1 of 1
73.53 (%)
Multiple-choice HAR QA (OpenTSLM, Langer et al. 2026), 8,222 test items; same-backbone Explicit-CoT 67.27, No-CoT 63.41; OpenTSLM-1B+Lapras 68.20, SLIP-1B+Lapras 70.75.
Sleep staging
opentslm-sleep-cot · Macro-F1 ↑
#1 of 1
79.49 (%)
Multiple-choice sleep-stage QA (OpenTSLM, Langer et al. 2026), 923 test items, 1-ch 1500-step windows; same-backbone No-CoT 78.05, Explicit-CoT 65.48.