SpatialClaw lands at NeurIPS and powers NVIDIA Nemotron 3.5 Super VL training data
CMHungSteven · x · 2026-10-16
The author announces SpatialClaw was accepted to NeurIPS 2026 and used to build agentic vision training data for NVIDIA Nemotron 3.5 Super VL, contributing 14,169 trajectories and 81,366 actions for video understanding. SpatialClaw is a training-free action-interface harness with no model-specific tuning; across 20 spatial benchmarks it lifts GPT-6 Astra from 71.3 to 77.4, Opus 5 from 58.5 to 73.5 (+15.0), GPT-6 Sol from 61.2 to 72.9, and Qwen3.8-27B from 59.8 to 67.3. The takeaway: even the strongest models gain significantly from a persistent Python kernel action layer. Paper, code, and report are public.
More from Research
- open-science: open-source local-first AI research workbench hits 5k GitHub stars — rohanpaul_ai · 2026-10-17
- Princeton/Penn's Recurrent Looped Transformer hits 100% on 256-bit parity extrapolation — TheTuringPost · 2026-10-17
- Princeton & UPenn propose Recurrent Looped Transformer: 100% parity at 256 bits after training on 40 — TheTuringPost · 2026-10-17
- Nature Medicine editorial: 800+ health systems deployed DeepSeek, but evidence lags AI rollout — EricTopol · 2026-10-17
- Open Math OS launches as open infrastructure for "Math 2.0" with free compute — LoubnaBenAllal1 · 2026-10-16
- Why Do Neural Networks Contain Legible Neurons? A Theory Finally Arrives — anshulkundaje · 2026-10-16