Ctrl+K
- claim-1-diffthinker-achieves-an-87-4-average-score-across-vision-centric-reasoning-tasks-outperforming-gpt-5-21-1-avg-by-314-2-table-1
- claim-2-diffthinker-outperforms-gemini-3-flash-41-3-avg-by-111-6-on-the-same-multimodal-reasoning-benchmark-table-1
- claim-3-diffthinker-outperforms-a-fine-tuned-qwen3-vl-32b-baseline-62-9-avg-by-39-0-table-1
- claim-4-diffthinker-reformulates-multimodal-reasoning-as-an-image-to-image-generative-task-using-flow-matching-rather-than-text-centric-chain-of-thought-section-3
- claim-5-diffthinker-s-inference-latency-is-reported-at-approximately-1-1-seconds-with-training-efficiency-comparable-to-sft-baselines-figure-5
- claim-6-ablations-show-performance-scales-consistently-with-training-data-size-and-peaks-at-a-classifier-free-guidance-scale-of-w-4-figures-8-9
- conclusion
- executive-summary
- 1.87 kB