Dimensions overall score 8.0

• Existing LVLM quantization methods typically measure token sensitivity at the modality level, which fails to capture complex cross-token interactions and falls short in quantitatively measuring quantization error at the token level.limitation 90%
• The proposed Quantization-aware Integrated Gradients (QIG) method leverages integrated gradients to quantitatively evaluate token sensitivity and pushes granularity from modality level to token level.method 90%
• Under 3-bit weight-only quantization, the QIG method improves the average accuracy of LLaVA-onevision-7B by 1.60%.result 95%
• The QIG method reduces the accuracy gap to full-precision counterpart to only 1.33% for LLaVA-onevision-7B under 3-bit weight-only quantization.result 95%
• The QIG method improves accuracy across multiple LVLMs and benchmarks with negligible latency overhead.result 85%
• As tokens interact within the model, the distinction between modalities gradually diminishes, suggesting the need for fine-grained calibration.technical 80%
• The QIG method reflects both inter-modality and intra-modality dynamics in token sensitivity evaluation.method 85%
• Large Vision Language Models have substantial computational and memory overhead that hinders practical deployment.limitation 90%

Competitive landscape

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

Startup potential card