Sensor-Language-Action Models (OpenSLA)

Yuekai Xu, Zitao Shuai, Yuzhe Yang (UCLA; the post credits Xu and Shuai as leads)

Summary

Action-necessity AUROC (%) is well above every baseline: supervised (ConvTrans, MMIM), time-series FMs (Chronos-2, Moirai, GluFormer), LLMs (GPT-5.6-Luna, DeepSeek-V3.2) and TS-LLMs (OpenTSLM, SensorLM).