Title: Design-MLLM: A Reinforcement Alignment Framework for Verifiable and Aesthetic Interior Design

Abstract:

The process of interior design involves translating requirements into visual plans, a task that demands the simultaneous satisfaction of verifiable spatial feasibility and comparative aesthetic standards. Although recent multimodal large language models (MLLMs) provide a unified foundation for interpreting user intent and generating design rationales, our empirical analysis highlights a significant contradiction in practical applications: MLLMs frequently generate layouts that are either unbuildable or aesthetically inconsistent. These observations suggest that merely incorporating domain-specific text is inadequate; instead, effective interior design necessitates an alignment mechanism that distinguishes hard constraints from soft preferences and coordinates them during the optimization phase.

To resolve this issue, we introduce Design-MLLM, a reinforcement alignment framework that optimizes a feasibility-first preference objective through a dual-branch, aesthetic-oriented reward system. Specifically, Design-MLLM (i) explicitly evaluates spatial feasibility via programmatic constraint checks, (ii) restricts aesthetic preference assessment to feasible candidates to prevent the selection of visually attractive but unexecutable shortcuts, and (iii) employs group-relative optimization to derive stable preference signals. Through this methodology, Design-MLLM acquires a controllable policy that reliably selects and generates solutions which are both executable and aesthetically coherent, moving away from the occasional production of visually appealing yet infeasible designs. Extensive experiments across various benchmark datasets confirm the superiority of Design-MLLM.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC