CEMIR vs. FISER: a comparative study of reasoning-based and enrichment-based approaches to pragmatic intent recognition

[arXiv] Ziya Hosseynzadeh, Farjad Fallah, Farnoosh Fallah and others conducted a comparative diagnostic study of two intent-recognition methods for human-computer interaction: FISER, based on structured social and embodied reasoning, and CEMIR, which achieves multimodal augmentation by synthesizing narratives and photorealistic images through text prompts. The authors tested four large models ranging from Qwen2.5-Omni-3B to GPT-5 on the HandMeThat v2 benchmark. Statistical analysis with multiple-comparison correction showed no statistically significant difference between the two on any model or difficulty level; all sixteen model-and-difficulty combination comparisons and ablation experiments targeting CEMIR's narrative and visual components showed this result. The paper states that performance is mainly related to backbone model capability and the inherent difficulty of ambiguity, and on this basis proposes a boundary condition for text-derived multimodal augmentation: when the relevant task constraints already exist in the text instruction, synthetic multimodal augmentation is not necessarily superior to structured reasoning over the same text.