Intent recognition is a fundamental problem in human‑robot interaction, requiring agents to infer the speaker's hidden goals to resolve ambiguous natural‑language commands. Recent work has focused on two paradigms: structured symbolic reasoning (e.g., FISER) and multimodal enrichment (e.g., CEMIR). The former applies social and embodied reasoning over the given text, while the latter synthesizes narratives and photorealistic images from prompts to supply extra semantic cues.
We benchmarked both approaches on the HandMeThat v2 dataset using four large language models ranging from Qwen2.5‑Omni‑3B to GPT‑5, covering four difficulty levels. After applying multiple‑comparison correction, none of the sixteen model‑by‑difficulty comparisons showed a statistically significant performance gap, and an ablation that removed CEMIR's narrative and visual components yielded the same pattern.
Analysis indicates that the backbone model capability and the intrinsic difficulty of the pragmatic ambiguity dominate results, rather than the presence of synthetic multimodal augmentation. When all relevant task constraints are already encoded in the textual instruction, text‑derived multimodal enrichment offers no measurable advantage over structured reasoning on the same text.
Review