Verified, then logged.
Sun 22 Nov 2026 · 00:17 KST / 15:17 UTC한국어 | ENSearch /

Breaking01 / AI

Spatial reasoning gains show limited generalisation

Researchers reported gains from spatial training questions alongside limits that varied by model.

AI-generated editorial illustration for this article

Researchers reported on arXiv that LLaVA-NeXT’s spatial accuracy rose by 23.1 percentage points. The gain was averaged across answer-option orders, but generalisation was limited.

The paper measured the improvement on CV-Bench, an image-understanding benchmark. Researchers further trained the model using hard negatives, or plausible but incorrect answers, and reported higher spatial-question scores after correcting the training loss, the measure used to guide learning.

According to the paper, the gains depended on spatial training questions. The researchers compared the results with control groups trained on other question types.

The paper also compared performance on CompHall, a benchmark for compositional reasoning—understanding how elements relate to one another. The researchers detected no improvement for Idefics2 on that benchmark.

The paper reported a stronger tendency for LLaVA-1.5 to answer yes in an object-hallucination evaluation, which tests whether a model reports objects that are absent.

The researchers found no evidence that failures in compositional reasoning caused hallucinations. Their findings left the reported spatial gains bounded by the training questions and the model tested.