CaST-Bench - QA Faithfulness: leaderboard

Evaluates causal chain-grounded spatio-temporal reasoning for video question answering on CaST-Bench: 2,066 questions over 1,015 videos whose causal chains are annotated as temporal segments with per-second bounding boxes. Predicted evidence instances are greedily matched one-to-one against ground-truth instances by coverage-aware spatio-temporal IoU (vIoU*), the temporal IoU multiplied by the mean per-second spatial IoU over the overlapping seconds. Faithful answer accuracy is the share of questions answered correctly and supported by visual evidence, requiring at least one matched evidence instance at vIoU* of 0.1 or above; a correct answer without such evidence counts as spurious instead, so faithful and spurious rates sum to QA accuracy. This is the open-submission Space leaderboard for CaST-Bench v1; published fractions are converted to percentages. This series reports CaST-Bench - QA Faithfulness using Faithfulness (%).

Metric: Faithfulness (%). Source: huggingface.co. Saturation forecast: Around February 2027. 7 models tracked.

Top models

No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.

Interactive version: theaggregate.ai/benchmark?slug=cast-bench-qa-faithfulness · How It Works · Data refreshed daily, snapshot 2026-11-05.