We argue that evaluating LLM agents in physical environments requires execution-based verification, the same standard we apply to code, with physics providing the ground truth. Construction tasks offer exactly this test.

We propose RecRec, which decouples reasoning from prediction in sequential recommendation. Following a TRM (Tiny Recursive Model) style, a recursive reasoner repeatedly updates an intermediate reasoning state and uses that state to refine the latent interests used for prediction. The number of reasoning steps can be adjusted at inference without retraining.

We introduce BuildArena, the first physics-aligned interactive benchmark for language-driven engineering construction, comprehensively evaluating LLMs' capabilities for construction automation under strict physical constraints through a highly customizable agentic engineering framework with static and dynamic mechanics tasks.

We investigate how far text-based collaborative filtering can go with extremely large language models, revealing key insights on leveraging language models for recommendation systems.