Sequential Context Engineering for Zero-Shot Recognition of Procedural Tasks in Egocentric AR

Abstract

We propose a sequential context engineering framework for zero-shot recognition of procedural tasks in egocentric AR. While large vision-language models excel at zero-shot image classification, they lack temporal reasoning required for procedural understanding, and fully fine-tuning them degrades generalization capabilities. The proposed method preserves frozen CLIP encoders while introducing lightweight modules: an intra-action attention module for temporal aggregation and an inter-action context module that leverages previous predictions to generate context-aware prompts. Evaluated on EGTEA Gaze+ dataset, our approach shows promising results with notable improvements over baselines while maintaining parameter efficiency. This work presents a step toward enabling temporal reasoning for VLMs in resource-constrained AR systems without sacrificing zero-shot generalization.