CockpitHAT: Hierarchical Attribution Framework for Embodied Multi-Agent Cockpits
The newly introduced framework, CockpitHAT, tackles the problem of Correctness Collapse in multi-agent systems using large language models (LLMs). This issue arises when high accuracy at the task level conceals significant failures at the process level, which is particularly dangerous in safety-critical environments like automotive cockpits. CockpitHAT substitutes positional windows with dependency-distance thresholds derived from interaction directed acyclic graphs (DAGs), incorporates multi-channel evidence through an embodied adapter, and enhances safety for high-risk failures during confidence-weighted analyst consensus. Additionally, the authors present CockpitBench, a benchmark comprising 212 annotated failure traces across dialogue, vehicle-state, environmental, and memory channels, each evaluated for ISO 26262 ASIL severity by three experts. This research is documented in arXiv:2608.01805v1, recently published.
Key facts
- CockpitHAT is a hierarchical attribution framework for embodied multi-agent cockpits.
- It addresses Correctness Collapse in LLM multi-agent systems.
- It uses dependency-distance thresholds from interaction DAGs.
- It integrates multi-channel evidence via an embodied adapter.
- It applies a safety-uplift to high-risk failures during confidence-weighted analyst consensus.
- CockpitBench is a benchmark of 212 annotated failure traces.
- Failure traces span dialogue, vehicle-state, environmental, and memory channels.
- Each trace is labeled with ISO 26262 ASIL severity via three-expert annotation.
Entities
Institutions
- arXiv