Explaining an artwork requires knowledge beyond visual recognition, including facts about its creator, subject, material, style, collection, and historical context. Knowledge graphs can provide such evidence, but artwork KGs are incomplete, especially for newly encountered artworks. We introduce ArtKGC, a leakage-controlled multimodal benchmark and evaluation protocol for artwork knowledge graph completion. ArtKGC links artworks to Wikidata facts, Wikipedia descriptions, and visual evidence while controlling ambiguous mappings, graph leakage, and descriptions that directly reveal held-out answers. We show that graph-only models are weak for unseen artworks, raw text can inflate evaluation through answer leakage, and visual evidence is strongly relation-dependent. As an interpretable diagnostic baseline, we use a validation-calibrated reranker that combines structural, sentence-redacted textual, and visual evidence. Experiments show improved completion in the primary inductive-artwork setting and modest downstream evidence gains for artwork QA and explanation diagnostics.

We introduce CARTBENCH, a museum-grounded benchmark for evaluating vision-language models (VLMs) on Chinese artworks beyond short-form recognition and QA. CARTBENCH comprises four subtasks: CURATORQA for evidence-grounded recognition and reasoning, CATALOGCAPTION for structured four-section expert-style appreciation, REINTERPRET for defensible reinterpretation with expert ratings, and CONNOISSEURPAIRS for original-copy discrimination under visually similar confounds. CARTBENCH is built by aligning image-bearing Palace Museum objects from Wikidata with authoritative catalog pages, spanning five art categories across multiple dynasties. Across evaluated VLMs, high overall CURATORQA accuracy can mask sharp drops on evidence-heavy and style-to-period questions; long-form appreciation remains far from expert references; and original-copy discrimination remains near chance. A held-out protocol-replication pilot on 120 artworks from five external institutions reproduces the main task-difficulty pattern, while not preserving the exact overall model ranking.

Current vision-language models (VLMs) perform well on many multimodal benchmarks, but still struggle with ARC-style abstract reasoning, where models must infer novel grid transformations from few examples. We analyze these failures through an information-processing decomposition into encoding, transformation, and construction, comparing three controlled settings: direct image-based solving, solving from model-produced text grids, and solving from gold text grids. Across existing ARC-1 and ARC-2 together with our controllable diagnostic benchmark V-ARAOC, we observe benchmark-dependent diagnostic signatures: ARC-1 is highly sensitive to encoding quality, ARC-2 remains difficult even after transcription noise is removed, and V-ARAOC is highly sensitive to constrained output construction. These results reveal benchmark-dependent information-processing bottlenecks in VLM-based ARC solving, with different benchmarks exhibiting distinct dominant failure modes.

Group bias in Large Language Models (LLMs) is a well-documented issue, its impact in high-stakes domains such as personalized nutritional advice remains under explored. This study introduces the USChainMains dataset to systematically evaluate LLMs, prompting them with names associated with specific racial and gender groups and rigorously quantifying the healthfulness of the generated meal recommendations against established dietary standards. The findings demonstrate that LLMs systematically recommend meals with significantly higher levels of adverse nutrients for names associated with Black, Hispanic, or male individuals, thereby reflecting and potentially reinforcing detrimental dietary stereotypes. Furthermore, our analysis of two common mitigation strategies reveals their limitations. While model scaling improves overall recommendation healthfulness, it is insufficient to eliminate the healthfulness gap between demographic groups. Notably, while augmented reasoning was effective in mitigating gender bias, it did not mitigate racial disparities. This work underscores the necessity of developing more nuanced, group-aware debiasing techniques to ensure AI-driven systems advance, rather than hinder, health equity.