ThinkBLOX: 3D Indoor Scene Generation with Progressive Reasoning

Traditional vision-language model (VLM) approaches to 3D scene generation typically follow a one-shot paradigm, where the entire layout is predicted in a single forward pass. While effective for simple scenes, this method struggles with interactive editing—such as moving or adding a single piece of furniture—because any change often requires a complete, computationally expensive re-optimization or reconstruction of the entire scene. Furthermore, one-shot models often lack the fine-grained spatial reasoning necessary to ensure that objects are placed in physically valid and semantically coherent positions.

ThinkBLOX addresses these limitations by reframing 3D scene generation as a progressive, state-conditioned reasoning process. Instead of generating a full layout at once, the model iteratively designs the scene step-by-step. At each step, the model generates a Chain-of-Thought (CoT) rationale that explains its spatial decision before outputting the specific coordinates and orientation for an object. This "reason-then-act" framework allows the model to maintain a consistent understanding of the scene's evolving state, making it highly effective for both initial scene generation and subsequent local rearrangements.

The authors introduce the ThinkBLOX-Data-200K dataset, which provides over 224,000 examples of procedural placement pairs, complete with multi-view context and CoT rationales. To refine the model, the researchers employ a two-stage training process: supervised fine-tuning (SFT) to establish basic reasoning skills, followed by a novel reinforcement learning (RL) scheme called Tier-Decoupled GDPO. This RL method organizes heterogeneous rewards—such as physical collision avoidance, semantic alignment, and reasoning consistency—into distinct tiers. By decoupling these rewards, the model avoids the common pitfalls of multi-objective optimization, leading to more stable and physically grounded scene layouts.