Evaluating Generative In-Vehicle Assistants: A Research Proposal for Mapping Claims to Human Outcomes

Abstract

Generative in-vehicle assistants are evaluated using heterogeneous measures, ranging from factual correctness and conversational preference to workload, sleepiness and driving performance. These measures support different claims. We propose a scoping review of how evaluation evidence supports claims about generative driver assistance, with particular attention to affect-aware support and the distinction between recognizing a state and improving it. An exploratory arXiv-led search has produced a candidate ledger and a provisional evidence map. This proposal specifies review questions, eligibility rules, extraction fields, quality considerations and a planned claim-to-measurement framework. Retrieval coverage and screening remain incomplete, and the framework has not been validated. No new driving experiments or intervention effects are reported. Planned outputs include an auditable review corpus, a measurement taxonomy, an evidence-gap map and evaluation recommendations bounded by participant role and study setting.

Keywords