IC-489State-of-the-art MLLMs fail at multi-step visual analogical reasoning, with best accuracy at 13% (Llama 3.2) on VOILA-WD and 29% (GPT-4o) on VOILA-ND, far below human performance of 71% and 70%

Nilay Yilmaz, Maitreya Patel, Yiran Lawrence Luo, Tejas Gokhale, Chitta Baral, Suren Jayasuriya, Yezhou Yang

SourceVOILA: Evaluation of MLLMs For Perceptual Understanding and Analogical Reasoning

Seven released MLLMs were evaluated on the VOILA benchmark, which requires models to infer a relational rule from a reference image pair and apply it to a new image pair to predict a fourth image. Performance degrades sharply across the four reasoning steps: models achieve 50-79% on image description, drop to 12-43% on relationship identification, and fall to 0.4-29% on applying the relationship. The gap between the best model and human participants is approximately 58 percentage points on VOILA-WD and 40 on VOILA-ND. The distraction rule in VOILA-WD further reduces accuracy for all models except Llama 3.2, which handles it better than other models.

Evidence
correlational
Key metric
Llama 3.2 applying relationship: 13.16% (VOILA-WD); GPT-4o applying relationship: 29.03% (VOILA-ND); Human: 71.36% (VOILA-WD), 69.69% (VOILA-ND); GPT-4o describing images: 78.94% (3i, VOILA-WD); GPT-4o identifying relations: 42.79% (3i, VOILA-WD)
Caveat
Evaluation uses GPT-4o as an automated judge with up to 10% error rate per step; image generation step evaluated only on GPT-4o, Seed-LLaMA, and Emu2 due to limited image generation capabilities of other models.
Model
GPT-4o, Llama 3.2 Llama 3.2-11B, Qwen2-VL Qwen2-VL-7B-Instruct, CogVLM2 CogVLM2-19B, Seed-LLaMA-8B, MolmoE-7B, Emu2 Emu2-37B
Concepts
Failure mode
Methods
Least-to-Most Prompting [primary]
Related work
Visalogy [context], VASR [context]
Related findings
IC-490, IC-491, IC-492
Extraction
automatic-extraction