aiqbench analysis
This research addresses "physhack," a failure mode where LLM-generated LEGO assemblies appear physically valid but lack geometric precision, proposing a data selection strategy that boosts performance using only 5% of standard training sets. The authors further introduce PVPO, a reinforcement learning method that couples physical-validity constraints with voxel-space rewards to significantly enhance the structural alignment of large language models in spatial reasoning tasks. These findings advance the frontier of sample-efficient fine-tuning for embodied AI agents capable of complex physical manipulation.
Sample-Efficient Post-Training for LEGO Spatial-Physics Reasoning
arXiv:2606.07602v2 Announce Type: replace-cross Abstract: LLM-based LEGO assembly requires both semantic grounding and physical feasibility. In this paper, we identify a data-induced failure mode, physhack, in which generated assemblies satisfy physical-validity constraints while remaining geometrically misaligned, semantically inconsistent, or poorly calibrated. To address this challenge, we propose a model-based data selection approach that uses only 5% of the original training data while improving semantic and structural alignment across multiple independent evaluators. We further introduce PVPO, a reinforcement learning method that couples physical-validity and voxel-space geometric rewards to strengthen the LEGO synthesis capabilities of LLMs. PVPO-trained DeepSeek-Llama-8B outperforms the frontier VLM Kimi-K3 across semantic, structural, and physical evaluation dimensions, even when Kimi-K3 is provided with a ground-truth image as an additional reference.
Abstract:LLM-based LEGO assembly requires both semantic grounding and physical feasibility. In this paper, we identify a data-induced failure mode, physhack, in which generated assemblies satisfy physical-validity constraints while remaining geometrically misaligned, semantically inconsistent, or poorly calibrated. To address this challenge, we propose a model-based data selection approach that uses only 5% of the original training data while improving semantic and structural alignment across multiple independent evaluators. We further introduce PVPO, a reinforcement learning method that couples physical-validity and voxel-space geometric rewards to strengthen the LEGO synthesis capabilities of LLMs. PVPO-trained DeepSeek-Llama-8B outperforms the frontier VLM Kimi-K3 across semantic, structural, and physical evaluation dimensions, even when Kimi-K3 is provided with a ground-truth image as an additional reference.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2606.07602 [cs.LG]
(or arXiv:2606.07602v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.07602 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 211 words
Related coverage
Sample-Efficient Post-Training for LEGO Spatial-Physics Reasoning
arXiv · Jun 9, 2026
The Machines that Make the Machines
NVIDIA · Oct 7, 2026
Develop Physical AI Reasoning, World, and Action Models with NVIDIA Cosmos 3
NVIDIA · Jun 1, 2026
Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk
Latent Space · Oct 8, 2026