Discovering Failure Modes in Vision-Language Models using RL

Vision-Language Models (VLMs) are increasingly used in safety-critical applications because of their broad reasoning capabilities and ability to generalize with minimal task-specific engineering. Despite these advantages, they can exhibit catastrophic failures in specific real-world situations, constituting failure modes. We introduce REVELIO, a framework for systematically uncovering interpretable failure modes in VLMs. We define a failure mode as a composition of interpretable, domain-relevant concepts-such as pedestrian proximity or adverse weather conditions-under which a target VLM consistently behaves incorrectly. Identifying such failures requires searching over an exponentially large discrete combinatorial space. To address this challenge, REVELIO combines two search procedures: a diversity-aware beam search that efficiently maps the failure landscape, and a Gaussian-process Thompson Sampling strategy that enables broader exploration of complex failure modes. We apply REVELIO to autonomous driving and indoor robotics domains, uncovering previously unreported vulnerabilities in state-of-the-art VLMs. In driving environments, the models often demonstrate weak spatial grounding and fail to account for major obstructions, leading to recommendations that would result in simulated crashes. In indoor robotics tasks, VLMs either miss safety hazards or behave excessively conservatively, producing false alarms and reducing operational efficiency. By identifying structured and interpretable failure modes, REVELIO offers actionable insights that can support targeted VLM safety improvements.

Vision-language models (VLMs) perform well on visual question answering with high-quality images but struggle when questions require knowledge beyond what is clearly and directly visible. In such settings, uncertainty quantification should not only indicate whether the model is likely to fail but also diagnose why it is uncertain, across dimensions such as perception, entity recognition, and knowledge retrieval. While prior work has focused on individual failure modes in isolation or treated incorrect answers as monolithic failures, we propose a unified framework for disentangling these failure modes and investigate whether pre-generation signals can predict these failure sources. Across a range of datasets and model families, we find a consistent pattern in VLM errors: some failures arise from visual or recognition bottlenecks, while others persist after the relevant entity is identified. Our main finding is that these failure sources can be predicted before decoding: recognition-related failures are best captured by visual-token representations, while failures that remain after recognition are better captured by prompt-conditioned hidden states. This pre-generation signal enables efficient failure-source prediction before the model produces an answer, allowing uncertain cases to be routed to targeted interventions such as image repair, entity recognition support, or external retrieval.

Vision-language models (VLMs) achieve high accuracy on many visual question-answering benchmarks, yet it remains unclear whether this accuracy reflects reliable use of the visual details a question depends on. We examine this tension through the lens of visual ignorance: cases where a model can represent task-relevant visual evidence without using it to determine the answer. Across three VLMs and twelve benchmarks, we track the answer-relevant information decodable at each decoder layer and measure when it becomes coupled to the final output. The visual evidence needed for the correct answer can already be read out from intermediate layers, yet it has little effect on the output until the final layers of the decoder. Causal residual interchange confirms that shifting the answer requires the late states, not the earlier ones. We then evaluate all twelve benchmarks under progressive Gaussian blur and find that substantial fractions of predictions survive the entire degradation sequence while continuing to receive benchmark credit. These results distinguish the availability of visual information from its influence on prediction, and show that standard accuracy can coexist with limited sensitivity to the tested visual detail. They motivate training and evaluation that make answer-relevant visual distinctions necessary to answer correctly.