ICML 2025 paper shows denoising models can raise spatial reasoning accuracy from under 1% to over 50%
CSProfKGD · x · 2026-07-27
Spatial Reasoning with Denoising Models (SRM)
The post shares an ICML 2025 paper from MPII on Spatial Reasoning Models (SRMs), a framework for reasoning over sets of continuous variables with denoising generative models.
- SRMs infer unobserved continuous variables from observed ones.
- The paper argues that diffusion/flow-matching models can hallucinate on complex spatial distributions.
- It introduces benchmark tasks to measure reasoning quality and hallucination in generative models.
- A key finding is that the denoising network itself can predict the generation order.
- Using sequentialization and sampling strategy insights, the authors report improvements on specific reasoning tasks from <1% to >50%.
- The post links the paper, code, and benchmark resources.
More from Research
- ReactBench v1 ranks GPT-5.6 at about 53% and Opus 5 at 49% on realistic React tasks — aidenybai · 2026-07-28
- ReactBench targets coding agents with real React tasks beyond unit tests — andrew_n_carr · 2026-07-28
- fchollet on iLands: Open-Ended Survival Environments Are Key to Shaping AI Intelligence — fchollet · 2026-07-27
- LocateAnything-Data opens up 12M images and 785M spatial annotations for multimodal training — ZhidingYu · 2026-07-27
- NVIDIA releases Molt, a PyTorch-native framework for agentic RL — dair_ai · 2026-07-27
- AI One-Shots 25k Lines of Code to Simulate Black Hole Electrodynamics in 5 Minutes — kylekabasares · 2026-07-27