0views0forks
Both configurations tile the image identically and run identical frozen encoders, so they see exactly the same pixels. The only difference is whether the tile tokens are compressed 4× by a concatenating pixel shuffle before reaching Qwen3. Passing all tile tokens gains 7.126 TextVQA points (11.07 standard errors); compressing them 4× gains 1.122, inside the noise floor. The two differ by 6.004 points, 6.43 standard errors of the difference.
The gain is specific to reading text. On the three non-OCR benchmarks the all-tokens configuration moves at most 0.7383 points — a targeted effect, which is what a resolution explanation predicts and what a general capacity increase would not.
What this does and does not license. The compressed configuration differs in two ways, not one: token count, and projector input width (4,608 against 1,152, forced by the concatenating pixel shuffle). Its projector output is badly conditioned — 8.1× the total variance, 97.25% of it in a single direction, and effective dimensionality outside that direction collapsing to 7.436 against 19.124 with all tokens. So the supported claim is that this 4× compressor loses the benefit, not that compression must. A better-conditioned reducer is the open question, and it is the difference between “don't compress” and “don't compress this way.”
Tiling with full tokens closes 35.69% of the 19.964-point TextVQA gap to Qwen3-VL-8B. Roughly 12.8 points remain, and they are not resolution: Qwen3-VL is trained on vastly more multimodal data at native dynamic resolution. Data scale is the next axis, not pixels.