Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• Visual attention in multimodal large language models exhibits pronounced inertia, remaining largely static once settled during early decoding steps.technical 95%
• Existing hallucination mitigation methods are inadequate for cognitive hallucinations that require inter-object relational deduction.limitation 90%
• Visual inertia is a key factor in cognitive hallucinations because attention to semantically critical regions remains persistently focused and fails to dynamically support relational inference.technical 90%
• The proposed Inertia-aware Visual Excitation (IVE) method breaks inertial patterns by modeling cognitive inference as the dynamic responsiveness of visual attention.method 95%
• IVE selects visual tokens that are dynamically emerging relative to historical attention trends while distinguishing tokens exhibiting inertial behavior.method 90%
• IVE introduces an inertia-aware penalty that discourages over-concentration and limits the persistence of attention within localized regions.method 90%
• IVE is effective across various base MLLMs and multiple hallucination benchmarks, particularly for cognitive hallucinations.result 85%
• The IVE method is training-free, requiring no additional model training.method 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card