Benchmark RadarAI BENCHMARK PROFILE
TRAPSBench
TRAPSBench is a procedurally generated video benchmark with 1,404 physics pairs to test epistemic restraint in VLMs, using Penalized Epistemic Calibration Score (PECS).
- Released
- 2026-08-13
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
It highlights that VLMs can internally detect when abstention is required but fail to express it, offering a metric for calibration and restraint evaluation.
Motivation
When visual evidence is occluded or chaotic, models should abstain.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.