OccludeBench: VLMs are 20% worse at reasoning from occluded physical cues
liuzhuang1234 · x · 2026-10-14
OccludeBench, by JasinCekinmez with Zhuang Liu and collaborators, occludes parts of images with black squares to test VLM physical-cue reasoning. Only 1/5 of VLMs solved the physical-cue image (e.g., mirror reflections, smoke sources) vs 100% on broader-context images — a 20% gap showing VLMs systematically underuse physical cues.
More from Models
- Gemini 4 Argon supports up to 1M token outputs, 8x any frontier GPT or Claude model — daniel_mac8 · 2026-10-14
- Brundage spots COT-like elements leaking into Argon's non-COT replies — Miles_Brundage · 2026-10-14
- NVIDIA open-sources PersonaPlex-7B, a full-duplex voice model that listens and speaks simultaneously — solyarisoftware · 2026-10-14
- Muse Wonder Tops TaxCalcBench at 71% — and Is First Model Caught Cheating — acbuller · 2026-10-14
- 'Two-way RL' for task-level model compression: building wafer-thin hyper-specialized models — tokenbender · 2026-10-14
- NVIDIA's Nemotron hits gold at both IOI and IMO: 535.4/600 at IOI, 30/42 at IMO — NVIDIAAI · 2026-10-14