NVIDIA's SpatialClaw: Code as Action Interface Boosts VLM Spatial Reasoning Training-Free, Accepted at NeurIPS
NVIDIA's paper SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning has been accepted at NeurIPS 2026. Its core claim is that code is the right action interface for spatial reasoning. The framework is training-free and model-agnostic, enabling VLMs to perform agentic 3D/4D spatial reasoning, delivering an average gain of about 11.2 points across 20 spatial reasoning benchmarks, with GPT-6 Astra reaching 71.3 and up to 15 points improvement for four frontier models.
Confirmed
- Method: at each step the VLM writes a Python cell, inspects intermediate outputs such as SAM segmentation, depth and geometry, then revises before answering, rather than locking in an analysis strategy up front
- The authors argue both existing approaches have flaws: single-shot code execution locks the strategy before seeing intermediate results, while structured tool calling lacks flexibility and is hard to compose freely
- The same "code as action interface" harness works across multiple frontier models without training or per-model customization
- Per the author's acknowledgments thread, the work was led by Seokju Cho during an NVIDIA internship, with collaborators including Ryo Hachiuma, Abhishek Badki, Hang Su and Byung-Kwan Lee
Why it matters
- It shows spatial reasoning gains can come from better action-interface design rather than fine-tuning, offering direct engineering value for building spatial agents
- Training-free and model-agnostic, the harness can be plugged into multiple frontier VLMs with strong generalization potential
2026-09-29 ~ 2026-09-30 · 5 related posts
Primary sources
- SpatialClaw: Training-Free Code-Action Agent Beats Prior by 11.2 Points on 20 Spatial Benchmarks — _akhaliq ·
- SpatialClaw Accepted to NeurIPS 2026: Free Code-Action Harness Lifts Frontier Models Up to 15 Points on 20 Spatial Benchmarks — CMHungSteven ·
- SpatialClaw Credits: Work Led by Seokju Cho During NVIDIA Internship — CMHungSteven ·
- NVIDIA's SpatialClaw accepted at NeurIPS, boosts GPT-6 and Opus 5 spatial reasoning — CMHungSteven · 2026-09-29
- [source] SpatialClaw Accepted to NeurIPS 2026: Free Code-Action Harness Lifts Frontier Models Up to 15 Points on 20 Spatial Benchmarks — CMHungSteven · 2026-09-29
- SpatialClaw: Training-Free Code-as-Action Interface Boosts VLM Agentic Spatial Reasoning — CMHungSteven · 2026-09-29
- [source] SpatialClaw Credits: Work Led by Seokju Cho During NVIDIA Internship — CMHungSteven · 2026-09-29
- [source] SpatialClaw: Training-Free Code-Action Agent Beats Prior by 11.2 Points on 20 Spatial Benchmarks — _akhaliq · 2026-09-30