2.5-D Decomposition for LLM-Based Spatial Construction

A neuro-symbolic pipeline that restricts large language models to two-dimensional planning has sharply reduced spatial errors in autonomous construction tasks, according to a paper posted to arXiv. The approach, called 2.5-D decomposition, lets a deterministic executor handle vertical placement, eliminating an entire class of coordinate mistakes [1][2]. The pipeline was tested on the Build What I Mean benchmark, which comprises 160 rounds of structure-building from natural-language instructions. GPT-4o-mini equipped with the pipeline achieved 94.6% mean structural accuracy across 12 independent runs, within 3.0 percentage points of the 97.6% ceiling imposed by architect-agent errors that no builder-side improvement can address [1][2]. By comparison, GPT-4o without the pipeline reached 90.3% and the best competing system reached 76.3% [1][2]. A controlled ablation attributed 50.7 percentage points of the accuracy gain to the 2.5-D decomposition itself [1][2]. Large language models are a type of machine learning model designed for natural language processing tasks and trained on vast amounts of text [7]. Their application to physical construction has been limited by systematic errors in generating three-dimensional coordinates. The researchers, led by Paul Whitten, addressed this by removing the vertical dimension from the model's output space entirely. The LLM plans block placements in the horizontal plane, while a deterministic executor computes vertical positions from column occupancy [1][2]. The pipeline transferred directly to edge hardware without prompt modifications. Nemotron-3 120B running locally on an NVIDIA Jetson Thor AGX matched the cloud result at 94.5% [1][2]. A transfer experiment on 500 IGLU collaborative building tasks confirmed the effect generalizes beyond the primary benchmark [1][2]. The paper states the principle applies to any autonomous construction or assembly task where gravity or other physical constraints fix one or more degrees of freedom [1][2]. The first version of the paper was submitted on 8 May 2026, with the latest revision dated 23 June 2026 [1][2].