New method reveals non-visual shortcuts in multimodal LLMs

Researchers propose a new method called Test-set Stress-Test (TsT) to identify exploitable non-visual shortcuts in multimodal LLMs. This approach involves training a text-only model on the test set's questions and answer options to reveal biases, rather than just evaluating a model's existing performance. The TsT method successfully identified significant biases in benchmarks like VSI-Bench and CV-Bench, where models could achieve high scores without processing images. However, on benchmarks like MMMU, the method found that high scores were more reflective of pre-trained knowledge than learnable test-set regularities. AI

IMPACT Highlights potential vulnerabilities in multimodal LLMs, urging for more robust evaluation methods to ensure genuine visual understanding.

RANK_REASON Academic paper proposing a new methodology for evaluating LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• CV-Bench
• Ellis Browne
• GPT-4o
• Iterative Bias Pruning
• Qwen2-7B
• Test-set Stress-Test
• VSI-Bench
• VSI-Bench-Debiased

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →