Oddly Satisfying Videos (OSVs) are a curious genre: repetitive motion, visual tactility, and ASMR-like sounds can trigger strong viewer reactions even though the clips carry few explicit emotional cues. That makes them a hard test for conventional emotion recognition systems, which usually rely on clear facial expressions or spoken sentiment.
The paper introduces PI-MLLM, a multimodal large language model that takes inspiration from psychophysics—the study of how physical stimuli translate into perceptual experience. By incorporating psychophysical priors into the model's reasoning, the authors aim to capture the subtle, sensory-driven emotional responses that OSVs evoke, rather than just parsing semantic content.
Because this is a single arXiv preprint, there are no conflicting sources to compare. The abstract does not yet provide quantitative results, so the claims rest on the proposed architecture and the authors' framing of the problem. Still, the work points toward a broader question: how should AI models handle emotions that are felt rather than stated?