Simulating Validity: Modal Decoupling in MLLM Generated Feedback on Science Drawings
The episode discusses 'Simulating Validity: Modal Decoupling in MLLM Generated Feedback on Science Drawings,' examining how multimodal models struggle to connect visual input (drawings) to textual feedback. Hosts analyze the paper's findings, noting that models often generate inaccurate or misleading feedback, even when trying to assess student work.