Dissecting Dissonance: Benchmarking Large Multimodal Models Against Self-Contradictory Instructions
This addresses a challenge for language beginners and vulnerable populations in multimodal interactions, representing an incremental improvement with a new benchmark and prompting technique.
The authors tackled the problem of large multimodal models (LMMs) failing to recognize self-contradictory instructions, introducing a benchmark with 20,000 conflicts and finding that current LMMs consistently struggle due to a lack of self-awareness, with a proposed method enhancing detection.
Large multimodal models (LMMs) excel in adhering to human instructions. However, self-contradictory instructions may arise due to the increasing trend of multimodal interaction and context length, which is challenging for language beginners and vulnerable populations. We introduce the Self-Contradictory Instructions benchmark to evaluate the capability of LMMs in recognizing conflicting commands. It comprises 20,000 conflicts, evenly distributed between language and vision paradigms. It is constructed by a novel automatic dataset creation framework, which expedites the process and enables us to encompass a wide range of instruction forms. Our comprehensive evaluation reveals current LMMs consistently struggle to identify multimodal instruction discordance due to a lack of self-awareness. Hence, we propose the Cognitive Awakening Prompting to inject cognition from external, largely enhancing dissonance detection. The dataset and code are here: https://selfcontradiction.github.io/.