SketchAgentBench: new benchmark tests 11 LLMs on 3D spatial reasoning from photos
wsxiaoys · reddit · 2026-11-07
Developer wsxiaoys built SketchAgentBench on top of a previously shared 3D coding playground, evaluating 11 LLMs on 3D spatial reasoning using 25 real photos.
- Task format: models reconstruct objects from a photo in a 3D scene, testing their ability to infer spatial structure and scale from 2D input
- Every result is inspectable in 3D at sketchagent.dev, with full session traces and editable scripts
- Example: leading model gemini-4-argon rebuilt a Tiny Glade cottage in 29 tool calls
A reproducible, fully inspectable new benchmark for anyone tracking models' spatial understanding — a prerequisite skill for embodied AI.
More from Research
- Daily agent #19: an overnight sequencing QC agent for molecular biology workflows — nlarusstone · 2026-11-07
- Can we CRISPR herpes viruses out of our DNA? An open question — moultano · 2026-11-07
- Inductive Bio Wins 4th Straight OpenADMET Blind Challenge, Beating 270 Entrants — graceisford · 2026-11-07
- Researchers Propose "Inter-AI Morality" Agenda: How AI Systems Treat One Another — dioscuri · 2026-11-07
- DSReg: multi-lab preprint disentangles predictive representations without decoders — jiqizhixin · 2026-11-07
- Steering one GQA key/value group makes a model claim "I am a user" 40 of 40 times — Sauers_ · 2026-11-07