| Paper_ID,Title,File_Path,Topic,Key_Weakness_1,Key_Weakness_2,Key_Weakness_3,Evaluation_Issue,Robustness_Issue,Citation_Impact | |
| 1,"Interpreting Second-Order Effects of Neurons in CLIP",papers/GPDcvoFGOL.txt,"Multimodal Neuron Interpretation","Polysemantic neurons (multiple concepts per neuron)","Selective effects (< 2% of images)","Direct effects negligible; second-order needed","Ablation methods fundamentally inadequate","Adversarial vulnerability via spurious concept correlation","Attribution under polysemy" | |
| 2,"BlueSuffix: Reinforced Blue Teaming for VLMs",papers/wwVGZRnAYG.txt,"Cross-modal Robustness/Jailbreak Defense","Unimodal defenses ignore cross-modal coupling","Bimodal methods degrade benign performance","Cannot defend against universal adversarial perturbations","Single-modality evaluation insufficient; need cross-modal testing","50-70% ASR with combined image-text attacks","Cross-modal jailbreak vulnerability" | |
| 3,"EUCLID: Supercharging MLLMs with Synthetic Data",papers/x07rHuChwF.txt,"Low-Level Visual Perception in MLLMs","Language bias dominates (text > multimodal by 26.8-28.7%)","Geometric perception < 30% accuracy (PointLiesOnLine)","Curriculum learning necessary; standard finetuning fails","Current benchmarks don't assess geometric perception","Models fail despite billions of parameters; noise-sensitive","Language prior overrides visual features" | |
| 4,"Cognitive Capabilities of Generative AI",papers/TjuS86sQv8.txt,"Multimodal Benchmark Evaluation","Perceptual reasoning at 0.1-10th percentile","Profound visual reasoning inability across architectures","Systematic failure despite 99.5th percentile on verbal tasks","Single benchmark insufficient; age/size effects unclear","Visual perception failures universal to all VLMs","Fundamental visual-linguistic integration issue" | |
| 5,"FIOVA: Five-In-One Video Annotations Benchmark",papers/Zggz6seq6F.txt,"Video-Language Comprehension","Information omission (4-15x shorter descriptions)","Limited descriptive depth vs humans","Uniform strategies on ambiguous content (no flexibility)","Single annotator baseline misses 80%+ of content","Inflexible handling of multiple valid interpretations","Lossy compression of visual information" | |