Study Questions Effectiveness of Latent Space Reasoning in AI Vision Models

Researchers have published a diagnostic study on arXiv examining a specific but critical failure mode in multimodal large language models (MLLMs) applied to video understanding: the inability to recognize when no valid answer exists among the provided options. Across diverse models and benchmarks, the study found that MLLMs overwhelmingly choose plausible-sounding distractors rather than flagging the absence of a correct answer—a behavior that undermines the reliability of these systems in real-world deployments. The failure was tested under three conditions: multiple-choice questions augmented with a 'None of the Above' option, open-ended generation with an explicit detection instruction, and standard evaluation without any guidance. The problem was found to be more pronounced in temporal reasoning tasks and worsened as frame sampling density increased, suggesting that richer visual input does not compensate for this reasoning gap. Chain-of-thought prompting was explored as a mitigation strategy and did substantially improve detection rates, but performance remained unsatisfactory, indicating that prompt engineering alone is insufficient. The authors conclude that explicit absent-answer detection mechanisms need to be built into multimodal systems rather than addressed through prompting workarounds. The paper is currently under review.