InTU Framework Infers Implicit User Preferences for Embodied Agents
September 9, 2026
The InTU framework enables embodied agents to infer latent user preferences from limited demonstrations to guide planning for underspecified instructions. The study introduces the Preference-based Planning (PbP) benchmark, comprising 5,000 evaluation groups and 290 preferences across three complexity levels.
HOW THIS AFFECTS YOU
●
builderYou can use the PbP benchmark to evaluate how well your embodied agents handle ambiguous real-world instructions.
●
researcherThe new benchmark provides a systematic way to measure preference inference in multi-level task planning.