Benchmark RadarAI BENCHMARK PROFILE
EditCLEVR
EditCLEVR is a paired-scene intervention benchmark for object-centric representations, with before/after CLEVR renders and a known attribute change. Includes metrics like SGIA and Delta-SGIA for semantic faithfulness, with probe-free diagnostics.
- Released
- 2026-07-19
- Readiness
- Runnable
- Primary field
- General AI
Why it matters
Provides a direct test of whether per-object representations behave correctly under controlled semantic edits, addressing a gap in evaluating compositional faithfulness beyond segmentation or single-image prediction.
Motivation
Object-centric learning aims to represent scenes as objects whose properties can be reused in new combinations.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.