Multimodal Task Interference: A Benchmark and Analysis of History-Target Mismatch in Multimodal LLMs
This episode analyzes the paper 'Multimodal Task Interference,' which identifies how current AI models struggle with context contamination. The hosts discuss how models fail when shifting from text history to image targets, a failure worsened by long, mixed conversations. The conclusion is that building robust AI requires designing for predictable failure points and implementing meta-cognition.