Preprint reports emotion reasoning gains from targeted prompt changes

Researchers report in a preprint that targeted changes to agent prompts improved multimodal emotion reasoning without updating model weights. On the MER2025 dataset, the method raised F-score by 0.6-2.2 points across four models ranging from 2B to 38B parameters, including an increase from 52.3 to 54.5 on Qwen2.5-VL-7B. Separate agents examine visual, acoustic and textual inputs, then combine observations into an emotion assessment with supporting citations. A verification stage checks evidence and identifies which component caused an error, generating prompt changes accepted only after validation on a held-out set. The gains were measured against a frozen static pipeline on one dataset.