2.5-D Decomposition for LLM-Based Spatial Construction

Paul Whitten, Li-Jen Chen, Sharath Baddam · 2026

Read 2.5-D Decomposition for LLM-Based Spatial Construction by Paul Whitten, Li-Jen Chen, Sharath Baddam (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-07066.txt.

Description

Paul Whitten, Li-Jen Chen, and Sharath Baddam help language models build block structures from spoken or written instructions. Models make systematic coordinate errors when placing blocks in three dimensions. Their pipeline lets the model plan only the flat, horizontal layout, while a fixed executor computes all vertical placements from column heights.

On a building benchmark, a modest model with the pipeline reached high structural accuracy, beating a stronger model working alone. Removing the vertical dimension proved to be the main source of the gain. The same pipeline ran on edge hardware with even slightly better accuracy, and prompt changes cut response time sharply while keeping quality high.

Catalog record

CREATORPaul Whitten, Li-Jen Chen, Sharath Baddam
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.07066

Browse the full library · About the Tower of Babel