Research map / Vision-language and multimodal / CLIP and zero-shot
Kiki or Bouba? Sound Symbolism in Vision-and-Language Models
Morris Alper, Hadar Averbuch-Elor
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use zero-shot CLIP text embeddings to score controlled sharp- versus round-associated pseudowords | CLIP · 648 pseudowords | AUC (γ): 0.50 -> 0.77 | Improved |
| Generate images from controlled sharp- versus round-associated pseudowords and score their image embeddings | Stable Diffusion · 648 pseudowords | AUC (γ): 0.50 -> 0.74 | Improved |
| Use CLIP text embeddings rather than Stable Diffusion image generations | zero-shot pseudoword probing · 648 pseudowords | AUC (γ): 0.74 -> 0.77 | Mixed |
| Use the prompt “a picture of a <w> object” instead of the main prompt (ablation) | CLIP probing · 648 pseudowords | AUC (γ): 0.77 -> 0.83 | Improved |
| Use the prompt “a <w> item” instead of the main prompt (ablation) | CLIP probing · kiki–bouba | ΔPkb: 52% -> -35% | Hurt |
| Use the prompt “a <w> drawing” instead of the main prompt (ablation) | CLIP probing · 648 pseudowords | AUC (γ): 0.77 -> 0.84 | Improved |
| Probe the multilingual Kandinsky model with Finnish prompts | Kandinsky · 648 pseudowords | AUC (γ): 0.50 -> 0.69 | Improved |
| Probe the multilingual Kandinsky model with Indonesian prompts | Kandinsky · 648 pseudowords | AUC (γ): 0.50 -> 0.67 | Improved |
| Probe the multilingual Kandinsky model with Hungarian prompts | Kandinsky · 648 pseudowords | AUC (γ): 0.50 -> 0.60 | Improved |
| Probe the multilingual Kandinsky model with Lithuanian prompts | Kandinsky · 648 pseudowords | AUC (γ): 0.50 -> 0.73 | Improved |
| Use larger unimodal sentence-encoder models (SMPNet or SDRoBERTa) rather than smaller models (SMiniLM or SALBERT) | unimodal text encoders · ground-truth adjectives | AUC (φ): 0.82, 0.71 vs. 0.57, 0.54 | Mixed |
| Match the sharp and round pseudoword classes to generated images in a human forced-choice task | Stable Diffusion generations · random sharp/round pseudoword user study | accuracy: 55% | Improved |
Closest papers by meaning
- Do Language Models Associate Sound with Meaning? A Multimodal Study of Sound Symbolism (AAAI 2026 · similarity 0.91)
- Cross-modal Associations in Vision and Language Models: Revisiting the Bouba-Kiki Effect (NeurIPS 2025 · similarity 0.88)
- EgoSound: Benchmarking Sound Understanding in Egocentric Videos (CVPR 2026 · similarity 0.83)
- From Senses to Decisions: The Information Flow of Auditory and Visual Perception in Multimodal LLMs (NeurIPS 2026 · similarity 0.83)
- RAIL: Rethinking Auditory Intelligence in Large Audio-Language Models with a CHC-Grounded Benchmark (NeurIPS 2026 · similarity 0.83)
- AVHBench: A Cross-Modal Hallucination Benchmark for Audio-Visual Large Language Models (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
