Datasets:

Modalities:
Image
Libraries:
Datasets
License:

error

#2
by mint20262026 - opened

Yes, a 37.5% success rate is clearly low. The paper reports 82.9%, so at that level, we would normally expect around 6–7 successes out of 8 samples, not just 3.

More concerningly, the failures were not all near misses:

The model counted the dice total as 72, while the correct answer was 77.
It selected the wrong candidate in the coordinate task.
In the image-selection task, its reasoning said an image should not be selected, but its action clicked it anyway.
In the click-order task, it understood the correct sequence but located the wrong icon.
It also misjudged the direction and target position in the slider task.
At this point, the only conclusion we can reasonably draw is:

The official fine-tuned checkpoint combined with the publicly released evaluation code performs far below the level reported in the paper when run locally using F16 weights with llama.cpp.

However, this is not enough to conclude that the paper is fraudulent. The authors recommend using the original BF16 checkpoint with vLLM, whereas we are using an F16 GGUF conversion with llama.cpp. The paper also omits important details about coordinate handling and image preprocessing.

There is no need to immediately run all 1,600 test samples. A more practical approach would be:

Randomly test 5 samples from each category, for a total of 40 samples. This should take approximately 1.5–2.5 hours.
If the success rate remains around 30%–50%, it would strongly indicate a serious reproducibility problem with the publicly released version.
If it rises above 70%, we can expand the test to 20 samples per category.
The current result of 3 successes out of 8 samples is already a significant warning sign.

Sign up or log in to comment