Autex: A Reliability-First In-Car Voice Assistant Agent for the CAR-bench IJCAI-ECAI 2026 Challenge

Abstract

We present Autex, a reliability-first in-car voice assistant agent submitted to the CAR-bench Challenge at IJCAI-ECAI 2026. Autex combines a deterministic multi-layer reliability kernel with a configurable LLM verifier to address three core CAR-bench failure modes: hallucinating unavailable capabilities, acting on ambiguous user intent, and violating domain policies. Epistemic guards fire before any LLM call, reducing the consistency gap between Pass@3 (latent capability) and Pass∧3 (deployment-readiness). We validate on 369 official CAR-bench trials across four frontier models, identify hallucination missing tool response as the dominant systemic weakness (0–14.3% Pass∧3 across all models tested), and describe the design choices made to address it. All model and provider configuration is environment-variable-driven, satisfying the Track 2 auditability and reconfigurability requirements.