REVELIO framework reveals hidden failure modes in vision-language models

A team of researchers (Chaudhary et al.) has developed REVELIO, a systematic framework for revealing interpretable failure modes in Vision-Language Models (VLMs). VLMs like GPT-4V and Gemini are increasingly used in safety-critical applications like autonomous driving and robotics due to their broad reasoning abilities, but they can still fail catastrophically under specific conditions. REVELIO defines a failure mode as a composition of interpretable domain-relevant concepts (e.g., pedestrian proximity, adverse weather) that consistently trigger incorrect behavior. Because searching over all possible concept combinations is exponentially large, REVELIO employs two complementary search strategies: a diversity-aware beam search that efficiently maps the failure landscape, and a Gaussian-process Thompson Sampling approach for broader exploration of complex failure modes.

In experiments, REVELIO uncovered previously unreported vulnerabilities in state-of-the-art VLMs across two domains. In autonomous driving, models exhibited weak spatial grounding and failed to account for major obstructions, leading to recommendations that would cause simulated crashes. In indoor robotics tasks, VLMs either missed safety hazards (e.g., spills, obstacles) or behaved excessively conservatively, generating false alarms that reduced operational efficiency. By producing structured, interpretable failure modes, REVELIO gives developers actionable insights to harden VLMs against specific real-world scenarios—a critical step before deploying these models in safety-critical environments.