Researcher speculates spatial reasoning leap comes from Blender training data
yoavartzi · x · 2026-09-12
Yoav Artzi discusses a model's striking spatial reasoning result on REMAP, a new benchmark he's building with anneyouw (release delayed by NeurIPS reviewing). He hypothesizes the model "bombed" the benchmark's ceiling because it was trained on heavy Blender data, and notes even ASTRA doesn't capture what humans do in other scenarios.
An informal but first-hand researcher's hypothesis about the source of a frontier model's spatial reasoning jump.
More from Models
- Continued pretraining vs RAG: an accuracy and performance comparison on Qwen 3.5 4B — funJS · 2026-09-12
- DeepSeek App Quietly Adds Four Read-Aloud Voices, Fueling New TTS Model Rumors — testingcatalog · 2026-09-12
- Watch OpenAI's Astra effortlessly drive a browser, as users ask about phone use next — infoxiao · 2026-09-12
- Sakana's Fugu Max hits OpenRouter: multi-agent orchestration, 1M context at $2/$6 per 1M tokens — SakanaAILabs · 2026-09-12
- "Why did you nerf Astra?" Users report model degraded days after launch — Significant-Ad6970 · 2026-09-12
- Why do Sonnet 5 and Opus 5 feel worse than GLM 5.3? Distillation's limits, dissected — baseten · 2026-09-12