Princeton's SSoN Benchmark: Best Vision Model Scores 39.7% vs Human 83.3% on Spatial Consistency
jiqizhixin · x · 2026-11-23
A Princeton AI Lab-led team presents Same Space or Not (SSoN), an arXiv benchmark testing whether models can tell if separate images describe one consistent space. Each of 1,000 questions (from seven scene sources, six edit categories, human-validated) gives three views of an unchanged indoor scene plus one controlled edit; the model must find the contradictory view and explain its choice without a trusted reference. Humans hit 83.3% vs 39.7% for the strongest vision-language model, dropping to 34.9% when a valid visual explanation is required — separating supported spatial judgments from lucky correct picks. A chart compares GPT-6 Astra Pro, GPT-5.6 Sol Pro, Gemini 3.5 Flash, Claude Fable 5 and others on true-solve vs partly-guessed rates.
More from Models
- Tesla demos FSD v15 autonomously navigating around a tree blocking the road — ZeroStateReflex · 2026-11-23
- Uncensored 26.4B model runs on a 12GB Mac at 85 tok/s with 8 parallel agents via Locals — SumitGup · 2026-11-23
- Dev: 99% of people shouldn't self-host local LLMs, and he only does it for the hardware — gnukeith · 2026-11-23
- Theo says open models trail by 7.5 months; GLM-5.5 and Kimi-K3.5 may flip the story — scaling01 · 2026-11-23
- Baidu open-sources Unlimited OCR2, an OCR model that parses entire books in one request — abhishek__AI · 2026-11-23
- Opus 5.5 spots a missing factor of 2 in a 2015 paper's bound during math checking — xennygrimmato_ · 2026-11-23