EmoHarness: Boosting Multimodal Emotion Understanding with Meta-Harness Evolving
Audio-derived evidence drove most of EmoHarness’s gains; automated harness edits added no statistically significant improvement.
On 5,507 MME-Emotion items, the system improved task-macro emotion recognition by 7.6–13.4 points over a single video-only call. It keeps the student model frozen and organizes visual observations, transcripts, and prosody descriptions as evidence for comparing candidate emotions. A stronger model edits the surrounding harness offline using errors and execution traces, without seeing evaluation items. Those evolved edits added at most 0.95 points, and none of their gains survived correction for multiple comparisons.
HF Daily Papers' note
On 5,507 MME-Emotion items, the system improved task-macro emotion recognition by 7.6–13.4 points over a single video-only call. It keeps the student model frozen and organizes visual observations, transcripts, and prosody descriptions as evidence for comparing candidate emotions. A stronger model edits the surrounding harness offline using errors and execution traces, without seeing evaluation items. Those evolved edits added at most 0.95 points, and none of their gains survived correction for multiple comparisons.
HF Daily Papers' note
score 4