Reading, Not Thinking: Understanding and Bridging the Modality Gap When Text Becomes Pixels in Multimodal LLMs
Open paper
Benchmark
Model
Rank
Results
arithmetic-reasoning-on-gsm8k
Qwen3-VL-8B (LM-only, self-distilled)
–
Accuracy: 92.72
code-generation-on-humaneval
Qwen3-VL-8B (LM-only, self-distilled)
–
Pass@1: 82.32
Rank counts only results with a code link.