Is VLA Reasoning Faithful? Probing Safety of Chain-of-Causation
A recent study investigates the faithfulness of Vision-Language-Action (VLA) driving models. The research reveals significant issues with reasoning fidelity, with only 42.5% accuracy in matching scene reality. The findings highlight concerns about trajectory fragility and reasoning-action consistency in various scenarios.
- ▪The study analyzed 300 Alpamayo-R1-10B inferences across 100 diverse PhysicalAI-AV scenarios.
- ▪Overall reasoning fidelity in the models was found to be only 42.5%.
- ▪The models exhibited 97.7% trajectory fragility under mild visual perturbations.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.17268 (cs) [Submitted on 17 May 2026] Title:Is VLA Reasoning Faithful? Probing Safety of Chain-of-Causation Authors:Nicanor Mayumu, Xiaoheng Deng, Patrick Mukala View a PDF of the paper titled Is VLA Reasoning Faithful? Probing Safety of Chain-of-Causation, by Nicanor Mayumu and 2 other authors View PDF HTML (experimental) Abstract:We present the first systematic study of faithfulness in Vision-Language-Action (VLA) driving models, analyzing 300 Alpamayo-R1-10B inferences across 100 diverse PhysicalAI-AV scenarios.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
