Improving Proactive AI Assistance with Hierarchical Procedural Understanding
The paper introduces ProactiveCoach, a training and evaluation suite for assistants that decide when to guide, stay silent, and vary instruction detail.
Its data structures guidance at phase, step, and action levels so models can track both fine progress and broader task context. The benchmark tests timing, appropriateness, and whether a model adapts when the requested guidance level changes. Fine-tuned VLMs trained with the hierarchical setup beat fixed-granularity supervision by up to 9.6 percentage points, according to the abstract. A router-based adaptive system then outperformed an in-context adaptation baseline by 57.1 percentage points across four guidance-level transitions. Source: HF Daily Papers' note
Its data structures guidance at phase, step, and action levels so models can track both fine progress and broader task context. The benchmark tests timing, appropriateness, and whether a model adapts when the requested guidance level changes. Fine-tuned VLMs trained with the hierarchical setup beat fixed-granularity supervision by up to 9.6 percentage points, according to the abstract. A router-based adaptive system then outperformed an in-context adaptation baseline by 57.1 percentage points across four guidance-level transitions. Source: HF Daily Papers' note
score 4