Verified, then logged.
Sun 22 Nov 2026 · 00:23 KST / 15:23 UTC한국어 | ENSearch /

Breaking01 / AI

Sophon tuning improves spatial reasoning in open models

Researchers traced spatial reasoning failures to a bottleneck in how models route information into answers.

An oblique close-up of a monitor shows a flagged intermediate mathematical step followed immediately by a corrected continuation, with all characters illegible. AI-generated image

Researchers unveiled Sophon and reported improved spatial reasoning in 7B-class vision-language models after targeted tuning, in a paper abstract published on 13 Nov.

The abstract describes Sophon as a diagnostic tool that asks about spatial relationships using groups of differently coloured points. Its design removes object types, scenes and textures as clues to the answer.

In researcher-published benchmark figures, open-weight models, whose trained parameters are publicly available, score 47–52% accuracy on Sophon. Leading models score 77–83%, while humans reach 84–94%.

The researchers said they repaired the route through which spatial information reaches answers using low-rank adaptation (LoRA), a method that adjusts selected parts of a model. Their published benchmark results put the tuned 7B-class models at the leading models’ overall accuracy.

The researchers’ published benchmark figures put open-weight models at 85–94% accuracy on existing spatial tests. They interpreted the lower Sophon scores as evidence of a bottleneck in routing spatial information, rather than an absence of that information.

The abstract does not disclose each model’s accuracy after tuning.