PRISM-BENCH: Diagnosing Grounded and Faithful Multimodal Reasoning Beyond Final-Answer Accuracy

A new benchmark, PRISM-BENCH, finds that the best of 25 open and closed vision-language systems, including GPT-5 and Claude Opus 4.6, reaches only 66.0% accuracy on its 1,000 expert-curated questions, and trails humans by 18.4 points on the four categories with a human baseline. The benchmark's perception probe attributes roughly 85% of frontier-model failures to misperception, placing the bottleneck at grounding, while its trace-alignment diagnostic predicts answer correctness for all 25 systems tested (macro AUROC 0.66) and survives length control. The authors also introduce CORE, an efficiency-aware metric that reorders 24.2% of model pairs relative to accuracy and remains uncorrelated with response length, and report that recent frontier gains fall on recognition rather than spatial grounding.

arXiv:2610.36554v1 Announce Type: cross Abstract: Final-answer accuracy reports whether a vision-language model was right, but not whether its answer was grounded in the image, whether its stated reasoning reflects that evidence, or what the answer cost to produce. Those are the properties that decide whether a system can be deployed. We present PRISM-BENCH, a benchmark of 1,000 expert-curated questions spanning five under-represented reasoning categories, in which every item carries an idealized reasoning trace written by one human annotator and validated by another. The traces support two diagnostics that accuracy cannot provide. A perception probe, in which the model answers from a description it wrote itself, attributes roughly 85% of frontier-model failures to misperception and places the bottleneck at grounding. Trace alignment predicts answer correctness for all 25 systems tested macro AUROC 0.66 and survives length control, so it measures reasoning content and not verbosity. We further introduce CORE, an efficiency-aware metric that reorders 24.2% of model pairs relative to accuracy while remaining uncorrelated with response length. Across 25 open and closed systems, including GPT-5 and Claude Opus 4.6, the best accuracy is 66.0%. On the four categories with a human baseline the best system trails humans by 18.4 points, and recent frontier gains fall on recognition and not on spatial grounding.