visual-tokens

Efficient Vision-Language Model inference and LoRA fine-tuning on ScienceQA, with dynamic resolution, visual-token compression, latency analysis, and accuracy–efficiency trade-offs using Qwen3.5-0.8B.

First-Divergence Consequence Analysis for masked visual generators: theory, audited coupling/instrumentation, and reproducible quantization-shock experiments.

End-to-end break-even evaluation for visual-token reduction in multimodal inference