Disentangling Vision from Meaning: A Neuro-Symbolic Framework for Semiotically Grounded Design Criticism

• Xiaofei Zhou,
• Soohong Kim,
• Yan Chen,
• Kailin Zhang,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.007, Download Paper (Free):

• Design Criticism
• Multimodal AI
• Dual-Path Framework
• Progressive Fine-Tuning
• Semiotic Alignment

Abstract

Current Multimodal Large Language Models (MLLMs) often fail in design criticism due to "semantic smearing," a phenomenon where dominant linguistic priors suppress subtle visual cues, resulting in generic or hallucinated evaluations. To address this limitation, we introduce a semiotic-driven neuro-symbolic framework that disentangles visual veridicality from cultural interpretation. Building on Qwen2.5-VL, our architecture incorporates Symbolic-Gated Adapters (SGA) and a Design Semiotics Knowledge Graph (DSKG) to enforce a dialectical separation between syntactic feature extraction and semantic reasoning. Furthermore, we replace deterministic objectives with a KL-divergence connotative loss, enabling the model to learn the probabilistic polysemy inherent in design discourse rather than forcing binary labels. Extensive experiments demonstrate that this disentangled approach significantly outperforms standard fine-tuning baselines in critical thinking depth (CTD) and style identification confidence (SIC). By bridging the gap between statistical pattern matching and structured semiotic logic, this work establishes a transparent, cognitively aligned paradigm for high-fidelity computational design critique.

Statistics

Cite this article